Subtitle: Markets treated a 40–60% AI-name drawdown, open-source mix-shift, Meta “rent compute” headlines, China DUV, and a credit/CDS scare as demand destruction. Private telemetry and a physical/financing constraint imply a different causal chain: contracted compute under-earns spot; open source transfers margin into infrastructure; Anthropic’s S1 is the pace car; Grok is the SpaceX sleeve the public book still underweights.
Date: 24 August 2026
Prepared for: Institutional investor / internal research (AP)
Classification: Thematic research. Not a trade recommendation. Not a restatement of PM memos.
Primary sources: Gavin Baker with Patrick O’Shaughnessy, Invest Like the Best EP.485, published 4 August 2026 (YouTube); Gavin Baker with Jason Calacanis and David Sacks, All-In, published 14 August 2026 (YouTube). Desk copies: /workspace/pm-memos/2026-08-04-gavin-baker-invest-like-the-best-ep485.md; /workspace/pm-memos/2026-08-14-gavin-baker-all-in-podcast.md.
External checks: Attempted 24 August 2026. Status recorded in Appendix. Speaker-stated numbers remain attributed until independently verified.
How to read labels: Source = spoken in a transcript. External-confirmed / partial / contradicted / unverified = desk check as of 24 Aug 2026. Inference = this desk’s synthesis, not Baker’s words. Hearsay = on-show second-hand (ARR, MoM, banker leaks). ASR = auto-caption risk.
July priced a demand break that private telemetry did not print. Baker’s hunt for “a single negative quantitative metric” after a month he called “2022 in a month” — AI names down ~40–60% / 50–60% from highs (Source, ILB) — returned accelerating GPU availability, GPU retail/spot, DRAM spot, and token growth, plus July pie growth at Anthropic/OpenAI/Grok/OSS (Source, ILB + All-In). Index-level press documents a real SOX/AI-chip selloff (External-partial: SOX ~−21% in July / ~−29% peak-to-trough from 22 Jun; single-name 40–60% remains Baker-attributed, not uniformly confirmed). The tape’s causal story (Meta rent-compute, OSS crush, China DUV, credit) is not the same as the demand story.
The mechanical bridge is contract much less than spot, then reprice, then operating cash flow, then less modeled credit need. Neoclouds locked cheap offtake to finance GPUs; the installed base now under-earns a still-tight spot market (Source, ILB: mid-~$2/GPU-hour Blackwell anecdote hoping just under ~$4 about seven months later; inference cloud planning +100% at expiry). Consensus, in Baker’s telling, still monetizes Blackwell+Rubin gigawatts at Ampere rates. If install-base reprice and Rubin lighting continue, hyperscaler operating cash flow — already large and rising in reported prints — can fund more of the build and take modeled credit demand out of the system (Source, ILB: 28 to 32 reported / ~28 to 35 adj.; ~$700B credit-need removal if OCF goes ~$1.3–1.4T to ~$2T). External-partial: MSFT+META+AMZN TTM OCF as of 30 Jun 2026 is ~$475B and accelerating; the specific 28-to-32 / 28-to-35 path and $700B credit arithmetic are not independently reconstructed.
Open source is a margin transfer into infrastructure, not a flops destroyer — and that is the live debate with Calacanis, not a consensus. A token is a token: same flops, memory, watts (Source, Baker, ILB). Mix-shift from high-margin frontier inference (debated 80/90/95%) toward OSS (~30% gross example) takes dollars out of the model layer and, with elasticity and routers (Fireworks Nexus, Together, Modal, Baseten), drives more dollars into GPU-hours. Baker’s All-In quantification: frontier may keep 65–85% of economic value while OSS becomes ~80% of volume. Jason Calacanis’s contrary claim — GLM ~90% cheaper than Claude Opus, corporate America will crush frontier pricing — is a volume argument that Baker treats as value-preserving for orchestrators. Do not collapse the panel.
Anthropic’s S1 is the demand and profitability pace car; Grok is the SpaceX sleeve the public book still prices as a rounding error. Baker is not an Anthropic shareholder (Elon-competitor rule, Source, All-In). He treats the FT $2T October leak as banker theater, refuses the over on $1T YE27 ARR, and puts $400–500B “in the zone” with Sacks; Calacanis takes a slower triple. If Anthropic brakes because demand is gone, there is “not enough room between the cars” — pile-up. If it is passed by OpenAI/SpaceX/US OSS, that is share-shift, not a crash (Source, Baker). SpaceX is public in this timeline (External-confirmed: Nasdaq SPCX, 12 Jun 2026, ~$1.77T IPO). Baker’s ILB claim that fundamentals improved post-IPO (Grok 4.5, Cursor) plus All-In’s Grok 4.6 Pareto / 4.7-in-weeks claim is the missing sum-of-parts leg.
Binding left tails are regulation/energization, a true demand-brake, and a dark-GPU glut if residuals are written off $30–50/W assumptions — not “is AI real.” Baker names regulation as the #1 risk (ILB: NY moratorium; All-In: Texas audit, atoms not bits). Nvidia’s ~$500B AI-factory platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR (External-confirmed, 10 Aug 2026) plus residual-value support “up to 25% of an opportunity” convert GPU compute into a financeable asset class — and create the glut steel-man if floors assume Elon-style $30–50/W spot. CoreWeave A100 rented into 2029 (External-confirmed, Aug 2026 earnings) is one data point for long residual life, not a proof that every new watt is safe.
The 4 Aug ILB recording sits inside the drawdown. The 14 Aug All-In recording sits after Nvidia’s 10 Aug financing announcement, after Grok 4.6 (12 Aug), and on the day Reuters broke Silver Lake/Workday talks (WDAY +17% External-confirmed). Between those two dates the public tape added the residual/asset-class story that ILB had described qualitatively as a “credit wrapper with a revenue share.” Near-term information events: Grok 4.7 (Musk: 3–4 weeks from 12 Aug, implying early/mid-September, External-partial); Anthropic confidential S-1 / rumored October window (External-partial, FT via syndication); Nvidia FY Q2 on 26 Aug; hyperscaler OCF path on the next print cycle; Texas ERCOT audit through year-end; NY EO 62 clock. The research window is weeks for Grok/IPO optics, about six months for the OCF/credit bridge, YE26–YE27 for the ARR band, and 2027–29 for memory LTAs.
Baker’s working knowledge-work TAM is ~$25T (range cited $25–65T on All-In; a bank AI-institute head called $25T “way low”). Token share of spend at “pilled” companies 20–25% (Dylan Patel ~30%; heard 50%) implies ~$5T from growth and/or labor substitution at 20% (Source, ILB). This is rhetoric, not a TAM model. McKinsey Global Institute has used ~$6.1–7.9T annual gen-AI economic potential (2023) and, in later work, ~$28.7T global automation potential at full technical capability — same order of magnitude, different construct (External-partial; not a confirmation of $25T times 20% equals $5T). More underwritable near-term gravity: Anthropic YE26 run-rate band $100–120B (FT investor leak, External-partial; company last self-disclosed ~$47B in May; Reuters/Bloomberg 17 Aug put July-end run-rate at ~$65B, below Baker’s on-show “>$80B” hearsay); Baker/Sacks YE27 $400–500B zone versus $1T over; Nvidia platforms designed to mobilize >$500B third-party capital over time (External-confirmed, not Nvidia revenue); Elon ~6–8 GW next year at $300–400B capex versus ~$100B just raised (Sacks, All-In, unverified); food-chain sketch ~$100B revenue / $50B compute / $30B chips per GW (panel, All-In).
| Market’s July story | Baker’s causal story | Desk inference |
|---|---|---|
| AI names down 40–60% = demand destruction | Private GPU/DRAM/token metrics accelerated; no quantitative negative found | Tape is not telemetry until OCF and S1 print |
| Meta “renting compute” = excess capacity / CapEx cut | Didn’t cut CapEx; telemetry more aggressive; Muse 1.1; SpaceX sold training clusters at a premium to contracted rates | Renting can be a price tell (spot tight), not a volume tell |
| OSS mix-shift / token-index flatten = AI over | Same flops per token; margin leaves frontier, enters infra; routers cut user spend, raise GPU-hours | Track GPU-hours and dollar mix, not frontier ASP alone |
| China DUV = ASML/semi-cap regime change | Phase change and ~25 years behind can both be true; order-book hit may be years out | Headline risk is not 2026–27 earnings for the EUV monopoly |
| Credit/CDS blowout = debt-funded bubble | Real if debt is required; dissolves if OCF reprice funds the build | Credit is the only July fundamental Baker grants |
| Anthropic share loss = peak | Pie “ginormous”; share loss with absolute growth is the base case | Pre-commit to a demand-brake versus share-shift rubric |
| SpaceX = launch + Starlink | Grok + Cursor + fastest/lowest-cost >500 MW energizer; orbital optionality | Public book still incomplete on the AI sleeve |
| Nvidia multiple compressed because over-earning | Decade-low forward P/E; financeability + wrapper + LTA game | Valuation claim is External-partial (low-teens/high-teens forward prints in Aug 2026 press; “10-year low” not independently audited) |
Baker’s own invalidation list (ILB + All-In), not this desk’s invention: (1) hyperscaler/ecosystem OCF stops accelerating; (2) sustained dramatic GPU-price contraction / easy availability (“not a single person has too many”); (3) sum of Anthropic+OpenAI+Grok/Cursor+OSS plateaus without pie growth; (4) buildout becomes debt-dependent into widening spreads; (5) regulatory blockage of energization (NY-style cascade; Texas audit as a red-state tell); (6) Anthropic demand-brake (not share-shift) implies pile-up; (7) S1 shows subsidized tokens at scale; (8) residuals written off $30–50/W assumptions produce dark GPUs. Additional desk risks: Valley selection bias; Claude-herding / 6-week “capacitor cycles”; Baker’s disclosed SpaceX book-talking and Elon-competitor avoidance; ASR/hearsay density on ARR.
Research agenda (2–4 weeks): reconstruct the OCF bridge from 10-Qs; build a GPU spot/contract tape; classify Anthropic wobbles as demand versus share before the first soft print; read Nvidia residual term sheets against the 25% language; independent Grok 4.7 evals outside Cursor; NY/TX/pledge dashboard; memory LTA commentary.
Monitoring dashboard (six panes): (a) demand telemetry versus narrative; (b) contract/spot/OCF/credit bridge; (c) OSS volume versus frontier value mix; (d) regulation/energization; (e) Anthropic demand-brake versus share-shift; (f) GPU residual/glut. Indicators in Section 9.
A thematic on the AI complex after the July 2026 public-market air pocket: how value, cash flow, and risk migrate across model labs, open-source inference, hyperscaler operating cash flow, GPU financeability, memory allocations, and the physical/political constraint on energization. The organizing question is not “is AI real.” It is whether July priced demand destruction or a narrative-plus-credit scare sitting on top of a still-tight physical system, with a contract-roll cash-flow bridge that public models have not yet fully encoded.
Time horizon: weeks (Grok 4.7, IPO/S1 rumors) through a ~6-month OCF/credit window, through a YE26–YE27 ARR band, through the 2027–29 memory LTA game, with a 3–5 year sleeve for orbital compute and software-industry structure. Geography: US-centric, with China OSS/DUV and EU regulate-first as constraints, not as the core underwriting.
The AI complex, for this note, has six stacked layers:
Where it sits in macro/tech: this is an atoms-constrained, credit-sensitive infrastructure cycle with a software-margin overlay, not a pure software multiple story and not a 1999 eyeballs story. Baker’s internet-unwind analogy is reserved for the case where the build is debt-dependent into widening spreads. His base instinct is that OCF reprice plus residual platforms make that analogy the wrong historical rhyme — unless S1 or GPU-rental data rehabilitate it.
Public markets observe listed AI/semi prices, hyperscaler FCF (CapEx-heavy, often deteriorating), token-index mix, CDS, and Claude-summarized newsflow. Baker observes private lab ARR, inference-cloud GPU-hours, DRAM spot, GPU rental anecdotes, and Valley meetings. The wedge is visibility, not necessarily disagreement about physics. Inference: any process that uses NVDA/SOX as a sufficient statistic for AI demand will systematically misread pace-car health until Anthropic/OpenAI disclosure cadence improves.
Private demand accelerates (tokens, GPU-hours, DRAM, lab ARR). The public tape sells AI 40–60% on Meta-rent, OSS, China DUV, and credit. The contracted GPU install base sits far below spot. Contracts roll, so monetization rises even if spot eases from peaks. Hyperscaler operating cash flow accelerates. Modeled credit need shrinks; the credit scare can fade without demand breaking. Alongside that spine: OSS mix cuts frontier margin and raises infra GPU-hours (Baker); Nvidia wrapper plus PE platforms unlock gigawatts that OCF alone cannot (All-In); memory LTAs lock 2027–29 allocations; regulation and energization can delay the bridge even if ROI is fine.
Two talks, one chain, with an important ten-day update. On 4 Aug Baker described the wrapper qualitatively (“credit wrapper with a revenue share if GPU prices are above a floor… not classic vendor financing”). On 10 Aug Nvidia announced platforms designed to mobilize over $500B of third-party capital with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, and Jensen’s essay framed “NVIDIA AI factory compute” as an investable asset class (External-confirmed). On 14 Aug Baker treated Wall Street’s CNBC validation as evidence that GPU compute is financable, named residual-value guarantees after about 3–4 years plus upside revenue share, and pointed at a Morgan Stanley note on capital-light royalties. Inference: the All-In hour is the financing-layer completion of the ILB hour, not a new demand thesis.
flowchart TD
A[Private lab + OSS + inference-cloud demand] --> B[Spot GPU-hours much greater than contracted rates]
B --> C[Contract rolls reprice the install base]
C --> D[Hyperscaler operating cash flow accelerates]
D --> E[Modeled credit need shrinks]
A --> F[OSS mix: volume up, frontier ASP pressure]
F --> G[Margin transfers into infra GPU-hours]
G --> B
H[Nvidia residual + PE AI-factory platforms] --> I[GW that OCF alone cannot fund]
I --> E
J[Memory LTAs 2027-29] --> K[Allocation lock-in for financeable GPUs]
L[NY moratorium / TX audit / EU regulate-first] --> M[Energization delay]
M --> N[Bridge timing slips; credit scare can return via duration]
O[Anthropic S1] --> P{Demand-brake or share-shift?}
P -->|Demand gone| Q[Pile-up: not enough room between cars]
P -->|Passed by OpenAI / Grok / OSS| R[Convoy continues]
S[Residuals set off 30-50 per W] --> T[Dark-GPU glut]
Frontier labs. Share is not the objective function while the pie is “ginormous” (Source, Baker, All-In). Anthropic can lose share on the margin to OpenAI, OSS, and Grok and still print historic ARR. Hubris is the cultural risk Baker flags (Dario “only private company in the world” — Baker takes the under; “SPF land”). Incentive at IPO: a responsible price that absorbs lockup so employees are not distracted; Baker reads a $2T print as evidence the roadshow cleared, maybe trading toward ~$3T chatter, and treats the leak itself as loser-of-lead-left theater.
Open-source and routers. Users want “slightly better outcomes at half the cost” (Source, ILB). AI natives RL their own OSS and route hard cases to a frontier checker. Public companies that burned 20× AI budget in three months use routers to stabilize reported spend while GPU-hours rise — a reconciliation trap for “AI opex slowing” bears (Inference, anchored to Baker). Jensen as “world’s largest supporter of open source” is aligned with Nvidia: OSS architectural diversity (Qwen/Kimi/DeepSeek/GLM diverging) favors flexible GPUs over fixed ASICs for residual life.
Hyperscalers. Under-earning versus spot until contracts roll. Meta “rent compute” is, in Baker’s read, not a CapEx cut: they did not cut CapEx; channel telemetry got more aggressive; they can sell or rent into a premium spot market. Microsoft’s “huge slug of capacity in June” was not in Q2 (Source, ILB) — a timing tell for the next print. Amazon sits in the Trainium LTA game. Combined OCF is the funding source that makes the internet-unwind analogy optional.
Nvidia. Three stacked advantages Baker emphasizes: most financeable GPU; power and land matchmaking; wrapper/residual plus revenue share that looks like a royalty on cloud, plus some equity stakes. Sacks’s gloss: “central bank of AI” via private markets; not Gurley-circular; asset-backed like aircraft financing; standardized reference designs needed to securitize. Inference: whoever sets residual floors and reference designs owns the securitization tollbooth. The incentive is to extend useful life (CUDA, fungibility) so financing rates fall — which is also how you get a glut if the life assumption is wrong.
Memory vendors. Tokens per unit of compute are memory-bound. LTAs (prepay, floor/ceiling) trade short-term upside for allocation durability. Breaking an LTA in 2027–29 risks losing future allocations when the cycle turns — “you’re out of business” (Source, ILB). Different from Apple versus Hynix/Micron monopsony. Baker’s CEO-of-Hynix thought experiment: copy Nvidia’s wrapper/surety plus revenue cut; Blackstone/Apollo likely suggesting variants.
Neoclouds / SpaceX compute. Contracted cheap offtake looks stupid mark-to-market until rolls. SpaceX’s spot-first, fastest/lowest-cost energizer status (the >500 MW/year club with hyperscalers, CoreWeave, Crusoe) is a demand-sponge tell: a vast compute dump “wasn’t even really a blip” (Source, ILB). NY hedge-fund short: spot −90% kills the revenue case. Sacks: Anthropic deal has 90-day cancellation both sides, so Elon sits near spot-minus-90 as a put.
Credit investors / PE. July CDS and Meta-adjacent bond pricing are “undeniable facts” (Source, ILB). Scary if debt finances the build. PE (Silver Lake/Workday) returning to battered software is, for Baker, a regime signal that OSS is a “godsend for the American software industry” rather than an extinction event — the world of 1–3 dominant frontier models is “a hard world for software”; OSS competitiveness re-opens the bid.
Regulators / local politics. Incentive is to be seen protecting household power bills, water, and jobs. Baker’s counter-incentives: behind-the-meter deals that “generally” lower local power prices; Trump pledge mechanics; ongoing plumber/electrician/HVAC jobs as “the best thing to happen for blue-collar wages in my lifetime”; water-usage book error he cites as 10,000× (ILB) / ~100,000× (All-In) — unverified, treat as his rhetoric. Industry “terrible job of PR.”
| Rank | Bottleneck | Binding? | Baker’s stance | Desk label |
|---|---|---|---|---|
| 1 | Energization / politics | Yes, on timing | Shortage is energy timing, not demand; turbines, diesel gensets, reconditioned airplane turbines; NY moratorium; Texas audit | Source; NY EO 62 and Abbott 3 Aug audit External-confirmed |
| 2 | Finance / residuals | Soft constraint Nvidia is trying to remove | Wrapper + PE platforms; glut if $30–50/W | Source; $500B platforms External-confirmed; residual 25% language External-partial |
| 3 | HBM / memory allocations | Structural through 2027–29 | LTA break = franchise risk | Source; LTA terms unverified |
| 4 | Credit spreads | Binding only if OCF fails | Only “real” July fundamental | Source; Meta CDS ~95bp, Sopaipilla $12.5B at 7.53% External-partial |
| 5 | Token demand | Not binding in his book | Pie ginormous; July accelerated | Source; private ARR hearsay |
| 6 | Open-source mix | False bottleneck if flops/token hold | Margin transfer | Source; Calacanis disagrees on value |
| 7 | Continual learning / SSI | Watch, not base | Temporary hit to training demand; training share already “very small”; SSI model “in August” | Source, ILB; scientific, not a trade |
This is the load-bearing mechanical claim. Long-term offtake was signed when the market expected old-GPU prices to fall in 2024/25. Instead, Baker says, old-GPU prices were “still going vertical in 2026” — a surprise to “anyone.” The contracted base therefore trades at a “massive discount” to spot. As those contracts expire, the new clearing price can be below today’s spot and still far above the old contract. That is how “spot could decline and compute will still get repriced higher,” answering ROI questions without requiring a perpetual squeeze (Source, ILB). Hypothesis (desk): this is also why FCF optics can stay ugly (CapEx still huge) while OCF accelerates (monetization of the stock of compute). The market’s favorite chart — semi cash flow up, hyperscale FCF down — “misses these private companies” (Anthropic, OpenAI, US OSS inference clouds).
The food chain as sketched on-show: Anthropic ~$100B revenue per GW, pay SpaceX ~$50B/GW, Nvidia ~$30B chips, then TSMC/Micron/SK Hynix. If the lead car (Anthropic) cuts the buy button because demand disappeared, trailing capex (SpaceX clusters, Nvidia systems, turbine orders, PE vehicles) has “not enough room between the cars.” If Anthropic is merely passed by OpenAI, Grok, or a US OSS champion, the convoy keeps moving. Inference: the highest-value forensic problem in the complex over the next two quarters is classifying any Anthropic “wobble” before the tape treats it as systemic. Pre-commit to the rubric in Section 9, pane e.
Zuck’s ~6,500-word manifesto (All-In chapter): open source, agent for everybody, superintelligence as invention not automation, safety via balance of power. Baker’s reduction: Anthropic/EA treat the technology as “too dangerous to distribute”; Zuck/Elon/Jensen treat it as “too dangerous to centralize”; “when given the choice it is always better to distribute.” If the US bans OSS or cartelizies two or three models (Sacks’s Biden-era “AEC-like” sketch), “we will lose AI… to China” (Source, Baker). Sacks’s irony: an FAA-for-AI that locked Anthropic’s lead would evaporate the lead and the pricing power. This is not decoration. It is a policy left-tail that sits next to NY/TX energization: centralize-to-protect-margins is, in this framework, simultaneously geopolitically and software-equity destructive.
TAM (rhetoric). Baker’s working number is ~$25T of knowledge work, with a $25–65T range on All-In and a bank AI-institute head calling $25T “way low,” pre-robotics. Token share of compensation at AI-pilled companies: 20–25% (his), ~30% (Dylan Patel, attributed), heard 50%. Arithmetic: $25T × 20% ≈ $5T of tokenized knowledge-work spend from growth and/or labor substitution (Source, ILB). Calacanis’s US-only sketch: 150–160M workers, salaries ~$10–12T, corps spend 5–10% of salary on tokens → ~$0.5–1T US AI spend (Source, Calacanis, All-In).
External check (24 Aug 2026): No reputable source was found that uses “~$25T knowledge-work TAM” as a standard figure. Adjacent constructs: McKinsey (2023) ~$6.1–7.9T annual economic benefits of generative AI; later MGI work on agents/robots cites ~$28.7T global economic potential at full technical automation capability and ~$2.9T US midpoint by 2030; Goldman Sachs ~300M jobs globally exposed, ~25% of US work hours. Status: Baker-attributed rhetoric; order-of-magnitude rhyme with MGI’s high-end automation potential; not confirmed as TAM. Do not put $5T in a model.
SAM (more underwritable, still hearsay-heavy).
| Object | Figure | Who | Status |
|---|---|---|---|
| Anthropic YE26 annualized run-rate | $100–120B | FT investors / All-In panel | External-partial. FT syndication confirms the leak: six backers, October window, $100–120B YE26, ~10×, 16–20× sales at $2T. Company last self-disclosed ~$47B (May). Reuters/Bloomberg 17 Aug: July-end run-rate ~$65B. Baker on-show “>$80B” is above the later Reuters print — hearsay, not confirmed. |
| Anthropic YE27 ARR zone | $400–500B | Baker + Sacks; Baker refuses $1T over | Source. Physical-limits argument, not a demand-disbelief argument. Unverified as a forecast. |
| Anthropic IPO mark | $2T leak; maybe prices $2T and trades ~$3T | FT / Baker | External-partial as a leak; not a company target (execs had not fixed valuation; quiet period after June confidential S-1). SpaceX IPO $1.77T (12 Jun) is the record the leak would break — External-confirmed. |
| OpenAI growth | 3–4×/yr historically; >20% MoM last 2 months | Sacks, hearsay | Unverified. Do not annualize two months of host hearsay into 10×. |
| Grok+Cursor ARR path | maybe ~$10B “pretty quickly” | Baker, ILB | Unverified. |
| SpaceX consensus “next year” | ~$73B; monetizing ~$50B/GW | Baker, ILB | Unverified. |
| Fund AI 8 GW ambition | “very implausible” / almost doesn’t believe | Baker on a Substack, ILB | Unverified; Baker himself is the skeptic. |
| Nvidia AI-factory capital | >$500B third-party, over time | Nvidia 10 Aug | External-confirmed. Not Nvidia revenue, not a single fund, not a commitment to one customer. |
| Elon GW/capex | ~6–8 GW next year = $300–400B vs ~$100B raised | Sacks, All-In | Unverified. |
| Food chain per GW | $100B rev / $50B compute / $30B chips | Panel, All-In | Unverified sketch. |
| Agentic users | ~250k–500k vs 7–8B people; already “acute compute shortage”; what happens at 1%? | Baker, ILB | Unverified; ASR on source. |
| Dwarkesh | 15× H100-year rental (~$250k) | Baker, ILB | Unverified as a market price; used as a probability-space shock. |
SOM (who can physically deliver). Baker’s >500 MW/year club: hyperscalers, CoreWeave, Crusoe, SpaceX — SpaceX “most… fastest… lowest cost” (Source, ILB). Physical SAM is turbine blades, interconnects, land, water politics, and residual-financeable GPUs, not “all knowledge work.” Baker/Sacks cluster on $400–500B Anthropic YE27 because of atoms, not because they think the TAM is small. That is the correct way to read the band: a feasibility SOM, not a demand haircut.
GPU-hour. Baker anecdotes (ILB): a “sexy” startup rented thousands of Blackwells at mid-~$2/GPU-hour; the same B200 cluster about seven months later hoping just under ~$4 (+50–60%); an inference cloud (he thinks Baseten) plans to pay 100% more for Blackwells at contract expiry. External-partial (24 Aug): CoreWeave list via aggregators: HGX B200 8-GPU node on-demand $68.80/hr (about $8.60/GPU-hr), spot $34.11/hr (about $4.26/GPU-hr); cheapest tracked B200 about $2.80/GPU-hr (Genesis Cloud). A May 2026 GPU-pulse note had B200 average ~$6.84/hr and rising. Public list prices in the ~$4–9/GPU-hr band for B200 are directionally consistent with Baker’s “hoping just under $4” spot-ish figure and make a mid-$2 contracted vintage look cheap. They do not confirm the specific startup anecdote or the +100% expiry plan. Spot, on-demand, reserved, and training-cluster contracted rates are not the same product.
Frontier versus OSS token margins. Baker (ILB): frontier inference margins debated 80/90/95%; OSS ~30% gross example. All-In: OSS makes frontier tokens more valuable as 250-IQ orchestrators of 150-IQ open models (Oppenheimer / farm-out analogy). Calacanis: GLM ~90% cheaper than Claude Opus. Inference: both can be partially true if one is talking user token price and the other orchestrator value share. The research task is to split token volume mix from revenue/profit mix. End-state Baker sketches: frontier 65–85% of economic value, OSS ~80% of volume.
Revenue per GW. Panel food chain $100B / $50B / $30B is a cartoon. Baker (ILB) separately says SpaceX is monetizing ~$50B/GW against consensus ~$73B next year for the company — mixing corporate revenue with $/GW compute tariffs. Do not combine them in a model. Hypothesis: if $50B/GW compute tariffs hold and Anthropic-class labs print $100B/GW of model revenue, the constraint is MW energized, not willingness to pay. That is the “acute shortage at 250k–500k agentic users” punchline.
Human versus model energy. Baker (All-In): approximating a brain ≈ power of ~1M US homes for 6–9 months; per-query ~1,000 homes. Used to argue that under compute scarcity, humans remain the most energy-efficient intelligences — i.e., labor substitution is not a free energy trade. Unverified physics-rhetoric; do not underwrite from it. Directionally it supports his “growth not mass displacement so far” labor read (more coder openings YoY; some new-grad pain; national accounts over ~18 months).
This is a capex-first complex. Amazon TTM FCF went negative (−$7.6B TTM Q2 2026) as PP&E, net, rose 64% YoY, “primarily… artificial intelligence” (External-confirmed, Amazon Q2 release, 30 Jul 2026). Meta Q2 FCF was $784M against $31.9B OCF and $30.1B PP&E (External-confirmed, Meta Q2, 29 Jul 2026). Microsoft FY26 additions to PP&E $116B versus $65B prior year; quarter ended 30 Jun additions $35.8B (External-confirmed, Microsoft FY Q4 2026 release). Inference: FCF deterioration is the accounting identity of the build, not proof of demand failure. Baker’s insistence on operating cash flow is the correct optical distinction.
The $1.3–1.4T → ~$2T “hyperscale OCF stock” line (Source, ILB) does not match a simple MSFT+META+AMZN TTM sum (~$475B as of 30 Jun 2026: Microsoft $182.9B, Meta $130.3B, Amazon $161.4B). Possible interpretations: a broader hyperscale set, a capitalized or run-rate construct, a GW×monetization model, or a misheard ASR figure. Status: Baker-attributed; not reconstructed. Put in Sources Needed. The ~$700B credit-demand removal is the delta in that same model if monetization moves from Ampere-like to discount-to-Blackwell. Also unreproduced.
What is reconstructed: the three names’ TTM OCF is large, growing, and coincident with collapsing FCF — the optical setup Baker is warning the desk not to misread. Microsoft net cash from operations $55.4B in the June quarter versus $42.6B prior year, $182.9B for the fiscal year versus $136.2B. Meta OCF $31.9B in Q2 versus $25.6B prior year, TTM $130.3B versus about $102B implied year-ago. Amazon TTM OCF $161.4B, up 33% YoY. Combined TTM ~$475B is not “28 to 32,” so the 28→32 / adj. 28→35 path remains partial: direction confirmed, vintage and adjustments (EU legal/fines) not independently tied to a public chart. Microsoft’s quarter did include discrete items (Anthropic investment gain $3.2B, VRP, Xbox charges) that affect earnings, not obviously the OCF add-back Baker described.
Nvidia’s >$500B platforms are designed to sit beside hyperscaler OCF. Sacks’s point is that Elon’s 6–8 GW / $300–400B need versus ~$100B raised is a finance constraint Nvidia is trying to nationalize into an asset class. Residual support “up to 25% of an opportunity,” project-by-project (External-partial, Huang), is the put that makes pension/insurance capital underwrite GPU life. Morgan Stanley, in secondary writeups, has modeled usage-linked revenue / royalty-like upside (External-partial; original note not retrieved 24 Aug). CreditSights’s counter in press: Nvidia is “writing a put,” pro-cyclical, cheap in the boom, expensive in a downturn.
Baker’s waves (ILB): AI natives not hiring humans, spending tokens; cutting-edge public companies; non-coastal US; Europe regulating first. Founder-led companies (ex-COVID over-hire): “nobody’s really laying people off” — bull case leans growth, not substitution. Cognition/Ramp/Stripe charts: highest AI spenders grow faster (he grants industry controls are imperfect). a16z/Iconiq work: gross profit per FTE elevated. Inference: if national accounts over ~18 months do not show a productivity pulse, the growth-not-substitution read weakens and the labor-politics left-tail (jobs narrative, moratoria) gets worse even if token demand remains fine.
Cursor ~15× efficiency (plan on frontier, execute on smaller) is a Source (ILB) microcosm of the OSS/router thesis: user outcomes up, frontier tokens used more sparingly, GPU-hours behind the router still rise. Grok 4.6 trained with Cursor workflow data (External-confirmed, model card 12 Aug); SpaceX completed the Cursor acquisition (External-confirmed, Cursor blog). Baker’s own caveat on All-In: Cursor benches may be “grading your own homework.” Wait on vibes (DHH-type) beyond benches.
Demand originates in knowledge-work tasks (coding first, then agents, then broader enterprise). It is satisfied by a mix of frontier API tokens, open-source weights run on inference clouds or private clusters, and application harnesses (Cursor, Claude Code, Grok Build, Harvey, Legora). Every token consumes flops, HBM, and watts. Watts require turbines, interconnects, land, and political permission. Flops require financeable accelerators and memory allocations. Financeable accelerators require residual assumptions that PE and banks will live with. Residual assumptions require a secondary market that still rents Ampere in 2029 and does not produce dark GPUs at $30–50/W build costs.
Value capture today is concentrated where scarcity and financeability coincide: Nvidia systems, HBM, energized megawatts, and (if Baker is right on orchestration) a subset of frontier tokens. Value is leaking, by design, from frontier list price into OSS volume and into the infra layer. Application-layer software that owns proprietary RL data (the Fireworks Nexus / Harvey / Cursor / Legora pattern) captures defensibility; generic wrappers do not. Hyperscalers capture the wedge between contracted GPU cost and spot-equivalent monetization as contracts roll — if they choose to price it. Neoclouds that signed cheap offtake transfer that wedge to their customers until expiry.
| Layer | Expand if… | Compress if… | Who said it |
|---|---|---|---|
| Frontier model APIs | Orchestration premium holds (65–85% of value); S1 shows cash generation | Jason’s 90% cheaper OSS wins on dollars; six-month lead evaporates via FAA-like licensing (Sacks) | Baker vs Calacanis vs Sacks |
| OSS inference clouds | Volume share → 80%; rule-of-40 growth with low burn | Frontier distillation / ASI path kills OSS role (Baker leans against) | Baker, ILB |
| Routers / harnesses | Multi-model default; 30–60% of tokens shift to own RL model | Single-model lock-in returns | Baker, ILB |
| Hyperscaler cloud | Contract rolls + Rubin lighting; OCF funds build | True CapEx cuts; FCF optics dominate cost of capital regardless of OCF | Baker, ILB |
| Nvidia | Wrapper royalties + residual platforms + OSS architectural diversity | Dark-GPU glut; residual put called; custom silicon wins allocations | Baker + Sacks |
| Memory | LTA floors hold through 2027–29; tokens/compute stay memory-bound | Severe sustained oversupply; LTA politics break | Baker, ILB |
| Power / turbines | Energization remains the bottleneck; capitalism ramps 24h shifts | Moratoria freeze queues; demand-brake cancels orders | Baker |
| Enterprise software | OSS keeps frontier from total app capture; PE bids (Workday) | 1–3 frontier models eat the application layer | Baker, All-In |
| SpaceX / xAI | Grok underwritten inside SPCX; compute dump absorbed | Spot −90% short lands; 8 GW miss plus energization failure | Baker; NY HF short as foil |
| Node | Role | Source locus |
|---|---|---|
| Anthropic / OpenAI / Grok+Cursor | Private demand “dark matter”; FCF/ARR acceleration; Anthropic pole then OpenAI/Grok back after compute race; S1 as pace car | ILB ~03:30–08:00, ~46:00–48:00; All-In ~02:36–27:32 and ~58:13–1:14 |
| Open source + inference clouds (Fireworks, Baseten, Together, Modal) | Mix-shift engine; token dark matter; multi-model routers; cash-efficient growth | ILB ~03:30–11:00, ~55:00–60:00; All-In Jason vs Baker |
| Hyperscalers (MSFT, META, AMZN; Google TPU; Amazon Trainium) | OCF bridge; CapEx telemetry; under-earning versus spot; LTA counterparties | ILB ~06:00–18:00, ~38:00–44:00 |
| Nvidia | Financeability apex; open-source ally; credit-wrapper + residual + revenue share; equity stakes; power matchmaking | ILB ~42:00–48:00; All-In ~58:13–1:14 |
| Memory (Hynix, Micron) + LTAs | Binding axis for tokens/compute; Game of Thrones allocations | ILB ~38:00–46:00 |
| Neoclouds / SpaceX / CoreWeave / Crusoe | Contracted versus spot; >500 MW club; SpaceX fastest/lowest-cost energizer | ILB ~05:00–08:00, ~70:00–75:00 |
| PE/banks (BX, KKR, GS, Apollo, BlackRock, Brookfield) | Validate GPU as financable asset class; underwrite offtake | All-In ~58:13; External-confirmed 10 Aug |
| Credit markets (Meta bond, CDS, real yields) | Only “real” July fundamental scare | ILB ~11:00–15:00; External-partial Sopaipilla/CDS |
| Regulators / NY / TX / pledge | #1 named risk; moratoria versus pledge/behind-the-meter truth gap | ILB ~62:00–68:00; All-In Texas audit |
| Power / turbines / gensets | Near-term soft bottleneck (energization), not demand destruction | ILB ~38:00–40:00; All-In CAT/Cummins/GEV/Siemens |
| SRAM / disaggregated inference | ROI uplift on installed+new base; older nodes; not everyone’s radar | ILB ~68:00–70:00 |
| Claude / public-equity herding | Homogenized Bayesian news interpretation; capacitor micro-cycle | ILB ~25:00–28:00 |
Accelerators. Nvidia is the incumbent with a financeability and software (CUDA) moat that Baker treats as more important than a given generation’s benchmark lead. Challengers: AMD, Google TPU, Amazon Trainium, startup ASICs (Etched; SRAM accelerators on older nodes). Baker’s 2027–29 LTA frame says custom silicon remains strategically necessary but pays a financing and memory-allocation tax versus Nvidia in scarcity regimes. Share is determined by pre-purchased supply chain, not only benches. Chinese OSS architectural divergence, in his All-In telling, favors flexible GPUs for residual life — a point against “ASICs eat Nvidia because OSS is China-shaped.”
Frontier labs. Anthropic is the current pace car on coding and, in Baker’s tea leaves, on cash generation. OpenAI is the share-taker via a coding pivot (Sacks) and, if Sacks’s >20% MoM hearsay is even directionally right, re-accelerating. Grok is the third on a cost-quality Pareto (Baker, All-In), still “small” at ~1–1.5T params, with 4.7 “significantly more capable” in a few weeks. Meta is the open-source T-Rex in Calacanis’s phrase; Baker’s softer pushback points at Jensen/Nemotron and a hyperscaler “double Lindy reverse” toward OSS (ASR; deal identity uncertain). SSI is a scientific watch (continual learning, August model, Nvidia “heavily involved”), not a current share threat.
Inference clouds. Fireworks (Lin as “absolute killer”), Baseten, Together, Modal: American OSS monetization layer, growing “almost as fast as the frontier labs” with little cash burn — “wild rule-of-40 optics” (Source, ILB). Not audited. Competitive risk is a pricing war that compresses the 30% gross example toward zero, which would break the “margin transfer into infra” story on the cloud P&L even if GPU-hours still rise.
Neocloud / compute landlords. CoreWeave, Crusoe, SpaceX, hyperscaler rental. SpaceX’s combination of speed-to-energize and 90-day-cancel flexibility (Sacks on the Anthropic contract) is a different product from a ten-year offtake. CoreWeave’s A100-to-2029 print is the residual-life existence proof the Nvidia financing story needs.
Steel-man Jason: GLM ~90% cheaper than Claude Opus; corporate America adopted open historically; Anthropic growth “going to slow considerably next year”; “no way they get to a trillion”; maybe triple to ~$300–400B. If he is right on dollars, Baker’s 65–85% frontier value share fails and the S1 profitability reveal is a last look at premium pricing.
Steel-man Baker: cheaper OSS tokens “massively inflate the value” of frontier orchestrator tokens (120 IQ cheap versus 160 IQ); volume is not value; Jensen would not be the world’s largest OSS supporter if OSS killed the GPU bill; inference clouds growing as fast as early labs with low burn is an infra tell, not a lab-death tell.
Desk inference: treat this as an empirical split to be measured, not a worldview. Pane (c) of the dashboard exists for this reason. A world where OSS is 80% of tokens and 80% of revenue is Jason’s world. A world where OSS is 80% of tokens and 15–35% of revenue is Baker’s world. Both are “OSS won volume.” They are opposite underwritings of Anthropic, Nvidia royalties, and software PE.
No recommendations. Mechanism of benefit is how the theme would transmit, not a call. Timing: early = weeks to two quarters (Grok 4.7, S1/IPO optics, next OCF prints); mid = six to eighteen months (contract rolls, YE27 ARR band, residual platforms seasoning); late = 2027–29 LTA/allocation and orbital optionality. “What must be true” is a hypothesis gate, not a price target.
| Company | Role in ecosystem | Mechanism of benefit | Key sensitivities / KPIs | Competitive risks | Timing | What must be true |
|---|---|---|---|---|---|---|
| Nvidia (NVDA) | Financeability apex; systems; residual guarantor; OSS ally (Nemotron) | Spot≫contract keeps GPU-hours bid; wrapper + residual + revenue share → royalty-like cloud; LTA game; PE platforms extend TAM | Forward P/E vs history; GPU spot/rental vs contract; residual floor vs $30–50/W; data-center GW; wrapper disclosure; FY Q2 26 Aug | Dark-GPU glut; residual put called; custom silicon + HBM allocations; circularity optics regardless of substance | Early (multiple, residual news) through late (LTA) | Demand telemetry stays tight; residuals conservative; OSS architectural diversity favors GPUs |
| Microsoft (MSFT) | Hyperscaler OCF; OpenAI exposure; Azure | Contract rolls + June capacity slug not in Q2; Copilot seats; OCF funds build | OCF not only FCF; Azure growth (43% in June qtr, External-confirmed); unusual items; CapEx vs telemetry; OpenAI investment P&L | True CapEx discipline; OpenAI share/terms; regulation of energization | Early–mid (OCF prints) | OCF keeps accelerating as Blackwell/Rubin light |
| Meta (META) | Hyperscaler OCF; Muse; open-source ideology; off-balance-sheet DC JVs (Beignet, Sopaipilla) | “Rent compute” as tightness not glut; CapEx telemetry aggressive; manifesto aligned with distribute-not-centralize | OCF $31.9B / FCF $0.8B Q2 (External-confirmed); CDS ~95bp; Sopaipilla 7.53%; CapEx guide vs channel; open-model quality vs GLM/Qwen | CDS/credit cost; political DC narrative; manifesto ≠ execution; Muse 1.1 overshadowed | Early (credit, CapEx optics) | Rent-compute is price discovery, not a CapEx-cut precursor |
| Amazon (AMZN) | Hyperscaler OCF; Trainium LTA counterparty; AWS | OCF +33% TTM to $161.4B; FCF negative on AI PP&E (External-confirmed); Trainium in memory game | TTM OCF vs FCF; AWS AI attach; Trainium allocation; not DSP/Mamdani (Calacanis/Sacks, out of Baker stack) | Custom-silicon mix; FCF optics; credit | Mid | Trainium does not lose LTA allocations; AWS monetizes tightness |
| Alphabet / Google TPU | Custom silicon; LTA counterparty; OSS adjacency | TPU as strategic necessity; LTA game | TPU share of internal+external; CDS (press: record); FCF | Financing/HBM tax vs Nvidia; China/OSS | Mid–late | TPU keeps allocations without breaking LTA politics |
| SK Hynix / Micron | HBM / DRAM; LTA counterparties | Memory dominates tokens/compute; DRAM spot accelerated in July (Baker demand tell); LTA floors | LTA commentary; HBM pricing/availability; DRAM spot; customer prepay; allocation shares | Severe oversupply; LTA breaks; Apple-like volume dynamics differ | Mid–late (2027–29) | LTAs stick even in an oversupply logic year |
| ASML | Lithography; China DUV foil | China DUV is a real phase change ~25 years behind EUV; order-book hit may be years out; market overreacted in July | China % of sales (~16% H1 2026 in press); indigenous DUV units (5 in 2026 / 20 in 2027 reported); EUV monopoly | Underestimating China speed; decoupling self-reinforcing; export-control step-function | Late as fundamental; early as headline | Learning-by-doing cannot be teleported (Baker) |
| CoreWeave | Neocloud; residual-life existence proof | A100 rented profitably into 2029 ≈ 9-year life from 2020 silicon (External-confirmed, 11 Aug earnings / Agrawal); H100 rebooked ~95% of original rate (press) | Prior-gen occupancy/rates; newer-gen mix; customer concentration; 90-day-style cancels | One data point; glut hits landlords; leverage | Early–mid | Prior-gen secondary stays bid, not liquidated |
| Crusoe | >500 MW club | Energization speed; behind-the-meter / gas adjacency | MW energized/year; offtake quality | Policy; fuel; execution | Mid | Stays in the >500 MW club |
| SpaceX (SPCX) / xAI / Grok / Cursor | Public (Jun 2026); compute landlord + frontier call; Cursor acquired | Grok 4.5/4.6 Pareto; 4.7 imminent; investors underweight Grok vs Starlink/orbital/terrestrial; fastest/lowest-cost energizer; supply dump absorbed with “no blip”; call (frontier) / put (rent compute, 90-day cancel) | MW/year; $/GW vs ~$50B claim; Grok+Cursor ARR vs ~$10B path; third-party evals (Databricks-class, not only Cursor); vibes (DHH); SPCX trading vs IPO $135 / ~$1.77T | Cursor-bench conflict (Baker flagged); NY short (spot −90%); 8 GW miss; disclosed boosterism (“talking our book”); energization | Early (4.7) through late (orbital) | Grok is a real third frontier lab, not a marketing sleeve; compute dumps keep getting absorbed |
| Anthropic (private → rumored IPO) | Pace-car demand + profitability reveal | Pie > share; S1 “breaks brains” if cash-generative; YE27 $400–500B zone, not $1T; $2T leak = banker theater | ARR vs $47B May / ~$65B Jul-end Reuters / panel $100–120B YE26 / Baker >$80B hearsay; FCF/token GM in S1; share vs OpenAI/OSS/Grok; IPO pricing vs lockup | Demand-brake pile-up; hubris; FAA-for-AI kills lead premium (Sacks); hearsay ARR; quiet period | Early (S1/October window) through mid (YE27 band) | Growth is pie, not a last look at premium; S1 not subsidized |
| OpenAI (private) | Share-taker; coding pivot; dead-man switch on capex (Calacanis) | If >20% MoM (Sacks hearsay) is real, pace-car is a pair; Baker: imminently cash-generative if not already | MoM growth unverified; coding KPIs; capex plan ($1.4T → ~$600B is Calacanis, unverified); SpaceX/other compute contracts | All growth figures on All-In are second-hand; 10× annualized from 2 months is not a model input | Early–mid | Coding pivot sustains dollar growth, not only share |
| Fireworks / Baseten / Together / Modal | US OSS inference + routers | Nexus-style 3-line integration; RL + router; growing nearly as fast as early labs, low burn; Baseten(?) +100% Blackwell at expiry = hyperscalers under-earn | Token volume; GPU-hour growth; burn multiple; router attach | Not audited; pricing wars; ASI distillation path | Early (telemetry) | GPU-hours rise even as user token prices fall |
| Harvey / Legora / Cognition | AI-native apps; router adopters; Cognition as dark-horse “Game of Thrones” | Proprietary RL data answers wrapper critique; Cognition index: high AI spenders grow faster (controls imperfect) | Routing mix 30–60% own-model; product adoption; index methodology | Selection bias; not public-market underwritable from these talks alone | Early–mid | Own-model share rises without flops/token collapse |
| Workday (WDAY) | Legacy enterprise software; PE bid as regime signal | Software “death spiral” ~12–18 months; Silver Lake talks; OSS godsend; PE sees oversold; world of 1–3 frontier models is hard for software | Deal confirmation (talks only, External-confirmed as Reuters talks, not a signed deal); +17% on 13 Aug (External-confirmed, close $206.45); peer PE bids; retention under AI | Talks fail; AI disruption still wins; Bending Spoons cost-cut is Calacanis color | Early (deal) / mid (wave or not) | OSS competitiveness is why PE bids, not a one-off |
| CAT / Cummins / GE Vernova / Siemens Energy | Turbines, gensets, jet-engine residual | Physical tip-of-spear; 24h shifts; capitalism ramps; Elon allegedly bought a turbine co personally (hearsay) | Backlog, lead times, utilization; interconnect queues; Texas audit; behind-the-meter share | Politics/moratoria; if AI capex slows, backlogs reverse | Mid | Energization remains the bottleneck, not demand |
| Etched (private; Baker investor context) | Custom silicon / SRAM-adjacent | Disaggregated inference (prefill / HBM attention / SRAM FFN) “really really positive for the ROI of AI” | Architecture vs HBM constraints; customer wins | Financing/HBM tax vs Nvidia; thin discussion; private marks | Late | Disaggregation ships and lifts ROI on installed base |
| TSMC | Foundry in the food chain | $30B/GW chips sketch includes TSMC | CoWoS/HBM packaging; utilization | China; customer concentration | Mid–late | Food-chain GW actually lights |
| Databricks (private) | Eval harness; not a thesis | $190B raise (External-confirmed, 13 Aug, $5B at $190B) used as sophisticated third-party bench for Grok vs “Fable 5” (ASR model name) | Eval methodology; $7B+ revenue run-rate (company) | One benchmark; not in Baker’s investment stack on-mic | Early (evals) | Benches are not Cursor-graded artifacts |
| Non-AI / traditional software (category) | Relative tape; OSS beneficiary | In the selloff month “everything but AI was vertical” (Source, ILB); OSS near frontier is a godsend for software cost/defensibility | AI vs ex-AI relative performance; software GM vs AI opex; multi-model adoption | If AI drawdown becomes broad risk-off, “vertical ex-AI” fails | Early (tape) | OSS keeps applications from being eaten |
Potential losers / disrupted incumbents (hypothesis, not a call). Pure frontier-margin bulls who ignored OSS elasticity. “AI is a debt bubble” books if OCF reprice works. NY-style SpaceX shorts if absorption continues. Under-financed custom-silicon stories in a scarcity LTA world. Enterprise software in a 1–3 model oligopoly without OSS competitiveness (the world Baker says is “hard for software”). Residual guarantors if floors were set off $30–50/W. Levered compute offtake structures in a demand-brake pile-up.
Names as color only, not table primaries. Dwarkesh (15× compute bull); TBU capacitor chart; Blackstone/Apollo wrapper suggestions; Benchmark/Starcloud/Eric (orbital sanity check); SSI (August continual-learning watch); Semianalysis-style token index; Fund AI Substack; Dylan Patel 30% token/comp; a16z/Iconiq GP$/FTE; Ramp/Stripe charts; Polymarket probabilities (Jason); “DART” $6B target (ASR identity, unverified); Boom (ad).
Each chain: primary observation → second-order → third-order → research relevance. At least two chains below are labeled contrarian.
Primary (Source, ILB ~05:00–18:00): contracted install base under-earns spot; +50–60% rental reprice anecdotes; OCF 28→35 adj.; $700B credit demand removable if monetization rises toward Blackwell. Second-order (Source + Inference): credit spreads and CDS can improve because of reprice, not only because Fed/macro softens. Baker: “ironically” credit metrics improve as the install base reprices — making credit easier even if they choose not to use it. Third-order (Inference): an equity market that sold AI on “debt-funded bubble” may be forced to re-rate the financing mix before it re-rates demand. Neoclouds that signed cheap offtake look stupid mark-to-market until rolls; hyperscalers look under-earning until then. Relevance: build a contract-roll / GPU-rental / hyperscaler OCF dashboard. Track Meta/MSFT/AMZN OCF versus FCF charts that omit private labs.
Primary (Source, ILB and All-In): frontier margins 80–95% versus OSS ~30%; same flops/token; elasticity; Jensen pro-OSS; routers cut user spend, raise GPU-hours; 65–85% / ~80% value/volume split. Second-order (Source): frontier may keep most economic value while OSS becomes most tokens; inference clouds print wild rule-of-40 growth with low burn; AI natives escape the “wrapper” critique via proprietary RL data. Third-order (Inference): public software that adopts multi-model routing may show AI opex deceleration in reported spend while underlying compute consumption rises — a reconciliation trap for “AI spend slowing” bears. Relevance: separate model-layer revenue from infra-layer GPU-hours in any deceleration claim.
Primary (Source, ILB ~38:00–46:00): memory dominates tokens/compute; LTAs stick even in oversupply logic; break = future allocation death; Nvidia most financeable. Second-order (Source): Nvidia credit-wrapper plus revenue share increases revenue/GW and competitive lock-in; memory vendors “should” copy; startups pay more for chips/HBM and finance worse. Third-order (Inference): custom silicon (Trainium/TPU/AMD/Etched) remains strategically necessary but faces a financing and memory-allocation tax versus Nvidia in scarcity regimes — share determined by pre-purchased supply chain, not only benchmarks. Relevance: diligence LTA commentary on memory earnings calls; Nvidia wrapper economics disclosure.
Primary (Source, ILB ~62:00–68:00; All-In Texas): regulation #1 risk; NY moratorium; water myth 10,000× / 100,000×; pledge lowers local power prices; blue-collar ongoing jobs. Second-order (Source): even deep-red pro-growth states say industry must tell its own story; ASCO/life-saving narratives unused; Texas audit shows the political constraint is not only blue-state. Third-order (Inference): if moratoria spread, OCF bridge timing slips even with spot greater than contract — credit scare returns via delayed energization, not via token demand. Political risk becomes a duration risk on the CapEx cycle. Relevance: track state/local moratoria, behind-the-meter deal terms, residential power-price differentials near DCs, pledge compliance. Leading indicator for GW slippage.
Primary (Source, ILB ~25:00–28:00): “Claude is Walter Cronkite for the stock market”; everyone feeds news to Claude/Claude Code; probabilistic sameness → Mauboussin diversity breakdown; TBU capacitor chart: entire cycle in six weeks that used to be a three-year cycle. Second-order (Inference): July can complete a full “cycle” of AI-adjacent equities before private ARR, DRAM spot, or GPU rental can falsify it. The technician worry Baker names — continuing decline without a clean villain except credit, “the bullet you don’t see” — is the herding signature. Third-order (contrarian inference): process edge accrues to desks that refuse Claude-summarized newsflow as a sufficient statistic and substitute primary telemetry (GPU-hours, OCF, rental indices, S1). The contrarian is not “be long because herding is dumb.” It is that false breakdowns and false V-bottoms (Liberation Day, DeepSeek) become the native volatility regime, so timing a “correct” fundamental can still force exits. Relevance: pair every narrative catalyst with a telemetry pane; do not let a six-week Claude cycle set the year-ahead scoreboard Baker invited.
Primary (Source, Baker, All-In ~1:35): software death spiral ~12–18 months; Silver Lake/Workday; OSS “godsend for the American software industry”; world of 1–3 dominant frontier models is “a hard world for software”; PE sees oversold. Second-order (Source + Inference): the consensus July software read is “AI eats applications, short the incumbents.” Baker’s read is that open frontier-adjacent models plus proprietary data is what lets American software companies exist at all. Workday +17% on talks (External-confirmed) is a regime signal, not a single-name event. Third-order (contrarian inference): the unexpected winner set includes legacy software that PE can take private and re-tool on OSS, and the unexpected loser set includes frontier labs if US policy cartelizies them into a 2–3 model FAA (Sacks’s irony) — because that world is both the one that kills software and the one Baker says loses to China. Policy that “protects Anthropic margins” is, in this frame, bad for Anthropic’s long-run lead, bad for US software, and bad for the China race. Relevance: watch PE take-private premiums in legacy software jointly with US open-model leadership (Meta/Nemotron) and export-control posture.
Primary (Source, All-In ~1:27–1:35; ILB ~70:00–75:00): Grok on Pareto; 4.7 imminent; investors ignore Grok; Anthropic $1B → ~$50B existence proof; ILB: fundamentals better since IPO. Second-order (Source, Sacks): call (frontier) / put (rent compute) with 90-day flexibility. Third-order (Inference): a successful Grok path increases correlation between “AI application” outcomes and SpaceX equity — and raises the opportunity cost of renting compute to Anthropic if spot rises. Orbital compute (Starcloud / Benchmark lasers, Starbase) is a real option that gets more real “every day,” not a 2026 cash-flow line. Relevance: underwrite SPCX with an explicit Grok scenario tree (fail → landlord; succeed → frontier lab), not Starlink-only. Wait on vibes beyond benches.
Primary (Source, All-In ~58:13–1:14): Nvidia matchmaker + residual guarantee + PE capital; Sacks “Fed of AI”; Elon 6–8 GW / $300–400B need. Second-order (Source, Baker): political friction caps overbuild; Ampere 9-year life supports residual; steel-man is glut at $30–50/W. Third-order (Inference): if residuals are set off telemetry that is too optimistic, the “asset-backed like aircraft” analogy fails the way MBS analogies fail — Baker himself invited the comparison as the worry to steel-man. Politics as accidental governor (moratoria insure against overbuild) is a bitter hedge: the left-tail that saves you from glut is the left-tail that delays the OCF bridge. Relevance: monitor GPU rental $/W, secondary prior-gen (Ampere/Hopper), PE fundraising pace, cancellation/renegotiation clauses (90-day Anthropic precedent).
Labor: Baker sees accelerating growth, not mass coder displacement (more openings YoY; new-grad pain). If that holds, token demand is complementary to headcount at founder-led firms and the bull case is growth. If national accounts over 18 months show substitution without productivity, the political jobs narrative that feeds moratoria gets worse. Water/power myths: whether the 10,000× book error is real or not (unverified), the narrative is already an EO in New York. Blue-collar RMA/upgrade jobs: Baker’s “best thing in my lifetime” claim is a political asset the industry has not used. China: DUV production start is External-confirmed as a report (Reuters/The Information, ~5 units 2026 / ~20 2027, Shanghai Aishengna); Baker’s “both ASML 2001 and new supply story can be true” is the right dual. Decoupling is “self-reinforcing”; Americans “lack clarity.”
Continual learning / sample-efficient learning (human ~20B tokens versus models ~300T; train on 10T then learn in the wild) is a temporary hit to training demand in Baker’s telling, with training’s share of compute already headed to “very small,” not zero. SSI model “in August,” Nvidia heavily involved (Source, ILB). He doubts it is net-negative for AI infra but stays open. Disconfirming: a working sample-efficient demo that cancels training clusters and (against his instinct) cuts inference GPU-hours. SRAM disaggregation failing to ship leaves ROI on installed base unimproved. China DUV plus learning-by-doing that is faster than his “~25 years behind” sketch would be a multi-year supplier-power shift, not a 2026 earnings event — still a left-tail for ASML/export-control geopolitics.
Valley selection bias: meetings, podcasts, “sexy startups” can miss enterprise ROI failure outside the coast. Europe “regulate first” and non-coastal lag might be the median, not the delayed wave. Agentic users at 250k–500k already in “acute shortage” is a strong claim; if 1% adoption never arrives because products do not cross from coding into broader knowledge work, the $25T rhetoric stays rhetoric. National-accounts 18-month test: no productivity pulse weakens the growth-not-substitution read and feeds the jobs/moratorium narrative.
Wrong OSS economics: if open models commoditize value not just volume, Baker’s 65–85% frontier value share fails — closer to Jason. Wrong pace car: if Anthropic is idiosyncratic (coding vertical peak) rather than systemic demand, reading the whole food chain off its S1 overstates beta. Custom silicon + HBM allocations that actually break Nvidia’s financeability premium. Grok benches that were Cursor-graded artifacts (Baker flagged the conflict himself).
NY EO 62 (14 Jul 2026, External-confirmed): up-to-one-year pause on ≥50 MW data centers while DPS writes a Generic EIS (energy, water, air, DAC impacts, noise); sales-tax exemption repeal pursued. Texas Abbott 3 Aug audit (External-confirmed): 474 GW interconnection queue, ~90% data centers, Batch Zero delayed, denial threat for non-disclosure; ERCOT completion targeted around 10 Dec 2026 in press. Trump Ratepayer Protection Pledge expanded late July (External-confirmed): voluntary, ~23 governors, ~55 utilities, White House claims ~80% of delivered power; Congress advancing H.R. 9340. Inference: the pledge is a political claim that DCs lower or at least do not raise household bills; Baker treats behind-the-meter economics as the truth under the claim. If pledge signers’ local residential rates still rise, the narrative Baker says the industry is losing gets worse in red states too. US FAA-for-AI / OSS restrictions: Baker’s lose-to-China condition; Sacks’s lead-premium-kill. EU regulate-first already in his wave list.
Wrong bridge: if contracted compute does not reprice higher (spot falls to meet contracts, or contracts were not as cheap as claimed), OCF acceleration and $700B credit relief do not arrive — July credit read was the correct fundamental. Buildout becomes debt-dependent into widening spreads — internet-style unwind, “very, very, very quickly” (Source, ILB). Off-balance-sheet DC JVs (Meta/BlackRock Sopaipilla $12.5B at 7.53%, Beignet $27.3B prior) are the credit market’s actual expression of the residual structure; weaker coverage on Sopaipilla (External-partial) is a leading tell that IG/project appetite is not infinite. Nvidia writing a put (CreditSights) is cheap until it is not.
July lacked a clean villain except credit; other narratives felt “ridiculous” to Baker; technician worry is the continuing decline you cannot see. Claude-herding plus capacitor micro-cycles produce six-week “cycles” that force process errors. Fidelity-friend advice he cites: navigate by the “dumbest, most superficial” narrative cycle; July risk-cutting chased stories that were “factually except for credit… not true.” Private marks (Anthropic $2T, Databricks $190B) driving public narrative without disclosure quality. Baker discloses SpaceX book-talking and Elon-competitor avoidance; Grok/SpaceX enthusiasm may be motivated. Sacks is a host with policy proximity — treat ideology segments as advocacy plus analysis.
Jason’s claim, stated as his not Baker’s: GLM ~90% cheaper than Claude Opus; corporate America will embrace OSS as it did historically; Anthropic growth slows considerably next year; no trillion; maybe triple to ~$300–400B. The implied chain is user price → enterprise RFP mix → frontier revenue deceleration → capex buy-button off → infra chain pile-up through the price mechanism, not through “AI isn’t useful.”
Where this is strongest: public-company AI budgets that already burned 20× in three months will route to cheaper tokens; that is also in Baker’s ILB. Where it is weakest inside Baker’s framework: user token price is not flops/token; routers can cut reported spend while GPU-hours rise; frontier orchestration value can rise as cheap 120-IQ tokens flood. Where it is weakest empirically as of 24 Aug: Reuters/Bloomberg July-end Anthropic run-rate ~$65B is still up from $47B in May and $9B at YE25; that is not a crush, though it is also not the panel’s $100–120B YE26. Update rule: if enterprise revenue mix (dollars, not tokens) shifts to OSS faster than orchestration premium appears, favor Jason’s branch. Pane (c).
The bear, which Baker himself steel-mans: GPU overbuild into $30–50/W assumptions (Elon expectation cited) creates dark GPUs the way dot-com created dark fiber; residual guarantees then look like MBS; S1 shows tokens were subsidized; Gurley-style circularity was the right model; Ampere-to-2029 is cherry-picked; political “governor” arrives too late. Calacanis’s dead-man switch: Anthropic/OpenAI buy button (OpenAI ambitions scaled $1.4T → ~$600B in his telling — unverified). CreditSights: Nvidia’s put is pro-cyclical.
Where this is strongest: residual support “up to 25% of an opportunity” is real language (External-partial); Sopaipilla priced wider with lower coverage; SOX/AI-chip selloff did happen; FCF is collapsing at AMZN/META while OCF rises — the classic late-cycle identity. Where it is weakest inside Baker’s framework: CoreWeave A100-to-2029 plus H100 rebooked at 95% is the opposite of liquidation (External-confirmed as a company claim); public B200 rental in the $4–9/GPU-hr band is not a crash; political headwinds ironically cap overbuild; tea leaves say Anthropic and OSS tokens are cash-generative — subsidy bears are “just wrong,” analogy: bearish because “oil is at $500 a barrel” when it isn’t. Update rule: GPU rental $/W cracks and prior-gen inventory rises and PE pauses AI-factory vehicles → elevate glut from steel-man to base. If S1 shows sustained negative unit economics, rehabilitate circularity/subsidy bears Baker dismisses.
Six panes, 14 indicators. Leading versus lagging noted. “If X, update thesis this way.”
| # | Indicator | L/L | If this happens… |
|---|---|---|---|
| 1 | GPU availability anecdotes + “too many GPUs” survey | Leading | Easy availability on a sustained basis → invalidate ILB demand hunt; move toward bear. Still-tight → hold reprice thesis. |
| 2 | DRAM spot / HBM availability | Leading | July acceleration was Baker’s demand tell. Spot down hard with rising inventory → demand or oversupply; separate from LTA politics. |
| 3 | Token-index volume versus mix (Semianalysis-style) | Leading | Mix shift to OSS with volume up → Baker infra-positive. Volume down across frontier and OSS → demand-brake. Mix shift with volume flat and GPU-hours down → Jason-plus-efficiency (flops/token broke). |
| 4 | SOX/NVDA/META tape versus Claude-newsflow | Lagging (tape) / leading (herding) | Another six-week “cycle” without telemetry change → herding, not information. Tape confirming telemetry → less variant. |
| # | Indicator | L/L | If this happens… |
|---|---|---|---|
| 5 | B200/H100 rental $/GPU-hr (spot, on-demand, contracted vintages separately) | Leading | Spot stays ≫ 2024–25 contracts into rolls → OCF bridge on track. Spot falls to old contract levels → bridge fails. Spot eases 20–30% but stays above old contracts → Baker’s “spot can decline and still reprice higher.” |
| 6 | MSFT+META+AMZN operating cash flow (adj. unusual items) | Lagging (print) / coincident | Next prints accelerate → 28→32 path was directionally right even if the vintage was opaque. Deceleration on a clean print → pause. |
| 7 | Meta/sector CDS, Sopaipilla/Beignet secondary, real yields | Leading | Spreads tighten as OCF rises → credit scare fading for the right reason. Spreads widen into rising issuance → debt-dependent path. |
| 8 | Nvidia residual term sheets / PE platform AUM actually raised (not “designed to mobilize”) | Leading | Real funds + conservative floors → finance constraint easing. MOUs only, or floors that assume $30–50/W → glut risk rising. |
| # | Indicator | L/L | If this happens… |
|---|---|---|---|
| 9 | Frontier versus OSS revenue mix (OpenRouter, enterprise RFP, inference-cloud commentary) | Leading | Tokens OSS, dollars frontier → Baker 65–85/80 split on track. Dollars OSS too → Jason. |
| 10 | Router attach / “burned 20× budget” public-co AI opex versus GPU-hour channel | Coincident | Opex down, GPU-hours up → reconciliation trap, infra still bid. Both down → demand. |
| # | Indicator | L/L | If this happens… |
|---|---|---|---|
| 11 | NY EO 62 clock; copycat state bills; Texas ERCOT audit outcome (target ~Dec 2026); pledge compliance versus local residential rates | Leading | Copycat moratoria + residential rates up near DCs → regulation becomes binding duration risk. Audit clears + behind-the-meter deals cut local rates → Baker PR-gap is still real but not binding. |
| 12 | Turbine/genset lead times and CAT/Cummins/GEV/Siemens backlog | Coincident | Lead times falling with orders still up → capitalism ramping as Baker expects. Backlog cancellations → demand-brake or finance stop. |
| # | Indicator | L/L | If this happens… |
|---|---|---|---|
| 13 | Classification rubric, pre-committed: (i) Anthropic ARR/usage down and OpenAI/Grok/OSS up = share-shift, convoy continues; (ii) Anthropic down and competitors down/flat = demand-brake, pile-up; (iii) Anthropic down on price with usage up = Jason mix, not pile-up; (iv) S1 FCF/token GM | Leading (usage) / lagging (S1) | Apply before the first soft print. Do not let the tape treat (i) or (iii) as (ii). S1 cash-generative → subsidy bears wrong. S1 subsidized → rehabilitate them. |
| # | Indicator | L/L | If this happens… |
|---|---|---|---|
| 14 | Prior-gen secondary (Ampere/Hopper occupancy and rates); 90-day cancel waves; PE pause; $/W on new builds versus residual floors | Leading | A100/H100 re-let near old rates (CoreWeave 95% anecdote class) → residual thesis holds. Inventory up, rates down, cancels up → glut, put gets closer to the money. New-build $/W at $30–50 with floors written there → steel-man becomes base. |
Speakers did not assign numeric probabilities. Weights below are analyst inference for research planning only, not Baker’s. Timeline anchors that are his or on-show: ~6-month financing/credit path; contract reprice over rolling periods; LTA game theory to 2027–29; Grok 4.7 in “a few weeks”; October Anthropic IPO window (FT/panel); YE26 $100–120B run-rate context; Baker/Sacks $400–500B YE27 zone; national accounts ~18 months; Ampere residual into 2029; “in a year” stock scoreboard; SSI August watch.
Analyst inference probability: roughly 30–35%. He is rhetorically close after failing to find quantitative negatives, but insists on humility (“maybe they are over-earning”) and names regulation as a live left-tail. Slightly lower than a pure-ILB bull because All-In adds the glut steel-man and the $1T refusal.
Assumptions (mostly Source): OCF keeps accelerating as Rubin lights and contracts roll; GPU rental stays firm or declines gently while still ≫ old contracts; open-source mix continues without flops/token collapse; credit spreads calm as modeled credit need fades; no cascading moratoria; Meta CapEx stays aggressive; Anthropic “soft decelerate” third-party data (the only soft negative he heard, ILB) proves noise; S1 confirms cash generation; YE27 ARR lands in or above the $400–500B zone without needing $1T; Nvidia residuals prove conservative; Grok 4.7 + vibes confirm benches; SpaceX/hyperscaler supply dumps keep getting absorbed; PE software bid broadens on the OSS thesis.
Winners (hypothesis): Nvidia (financeability + wrapper + OSS alignment); memory with sticky LTAs; inference clouds; AI-native software that owns RL data; power/turbine/energization chain; SpaceX if MW and $/GW and Grok print; selective enterprise software where PE sees oversold. Losers (hypothesis): “AI is a debt bubble” books; pure frontier-margin bulls who ignored OSS elasticity; NY-style spot-collapse SpaceX shorts if absorption continues; subsidy/circularity macro bears (Baker’s explicit foil); Claude-herded capacitor-style day-traders.
Regime-shift trigger into bull: next hyperscaler OCF print accelerates and GPU rental holds and S1 is cash-generative. Any two of three is not enough; all three closes the “maybe over-earning” humility loop.
Analyst inference probability: roughly 40–45%. Closest to Baker/Sacks spoken central tendency (under on $1T; $400–500B zone; physical limits; OSS good; pile-up only if demand dies) and to his lived dissonance: Valley more bullish than him, tape makes him feel foolish; technician “bullet you don’t see”; year-ahead scoreboard humility.
Assumptions: Demand telemetry stays positive but public multiples stay compressed because the market assumes over-earning; contract reprice helps OCF with lag; credit remains a recurring headline; China DUV / open-source scares recur and fade; some local regulatory friction without national freeze; Nvidia multiple stays “hard to understand” for longer than Valley likes; Anthropic grows very fast but not 10× again; Jason’s slowdown partially right on rate, wrong on “OSS kills frontier value”; Grok is real #3 but not instantly dominant; Workday is a one-off signal more than a wave immediately.
Winners (hypothesis): relative winners inside infra with financeability/memory leverage; multi-model software tools; selective neoclouds with roll timing; process edge from primary telemetry versus Claude consensus. Losers (hypothesis): anyone needing clean linear 10× extrapolation or clean linear collapse; pure “OSS zeros Anthropic” or pure “Anthropic only private company” maximalists (Baker rejects both); narrative-momentum sleeves that must “cut risk” every Claude summary; under-financed custom-silicon stories in a scarcity LTA world.
Regime-shift trigger into base from bull: S1 is fine but IPO prices well below $2T leak (headline risk Baker already framed as theater); Grok 4.7 slips; Texas audit delays GW without killing demand. Out of base into bull: residuals season with no dark inventory. Out of base into bear: see below.
Analyst inference probability: roughly 20–25%. Maps to his own invalidation list. Split internally (desk inference, not his): ~10–15% demand-brake; ~5–10% glut-at-wrong-residuals; ~5% regulation-duration that slips the bridge until credit re-bites. These overlap.
Assumptions: Anthropic (and peers) cut buy buttons because demand is missing — not because SpaceX/OpenAI/OSS took share; and/or GPU overbuild into $30–50/W assumptions creates dark GPUs; and/or S1 shows tokens were more subsidized than Baker’s tea leaves; and/or OCF stops accelerating; and/or sustained dramatic GPU-price crash and easy availability; and/or regulatory moratoria block energization so GW slip; and/or continual learning discontinuously crushes training and infra demand; and/or Grok benches fail to translate to vibes and the landlord put reprices down together.
Winners (hypothesis): short-duration quality; credits that were underwriting the scare correctly; skeptics who underwrote subsidy if S1 proves them right; non-AI “vertical” survivors of the July rotation if risk-off broadens (path-dependent — he noted non-AI vertical in the selloff month). Losers (hypothesis): highly leveraged neoclouds; memory/GPU names priced for perpetual scarcity; residual guarantors if floors were set wrong; SpaceX if both frontier call and landlord put reprice down; custom silicon without allocations.
Regime-shift trigger into bear: indicator 13(ii) (demand-brake, not share-shift) or indicator 6 deceleration or indicator 14 glut cluster. One is enough to reduce confidence in the YE27 zone; two is a regime shift.
Anthropic losing share while ARR rises. OSS token mix rising. Spot easing 20% from peaks while remaining above 2024–25 contracts. $2T IPO leak slipping or pricing below leak. China DUV headlines recurring. A six-week SOX cycle. Meta renting compute. Workday talks failing. Any of these can be loud. None, alone, is Baker’s pile-up condition.
Stop-research criteria (put these on the calendar, not in a drawer): (a) next MSFT/META/AMZN OCF print is cleanly decelerating — pause the reprice workstream; (b) GPU rental indices show a sustained crash and “too many GPUs” becomes the modal channel check — demand thesis break; (c) S1 (or a leak of audited figures) shows subsidized tokens — rehabilitate subsidy bears and re-underwrite the financing chain; (d) Anthropic usage down with competitors down — pile-up protocol, stop treating this as a share-shift thematic. If none of (a)–(d) hit in four weeks, the workstream continues into the six-month OCF/credit window.
P1 — must-do before any allocation-research decision
Hyperscaler OCF bridge. Reproduce Baker’s 28→32 reported and ~28→35 unusual-item-adjusted. Identify the EU legal/fine add-backs. Desk has TTM ~$475B (MSFT $182.9B + META $130.3B + AMZN $161.4B as of 30 Jun 2026) and directionally accelerating OCF with collapsing FCF; the vintage of “28 to 32” is still opaque. Data: 10-Qs, cash-flow statements, earnings slides. Expert: internet/hyperscaler analyst.
GPU spot versus contract tape. Verify mid-$2 → just-under-$4 /GPU-hour B200 anecdote class and “100% more at expiry” inference-cloud claims with channel checks (neoclouds, brokers, SemiAnalysis-type rental indices). Public list prices ($4.26 spot / $8.60 on-demand CoreWeave B200) are a start, not a contracted-vintage tape. Without this, the entire reprice→OCF→credit thesis is anecdote. Expert: AI-infra channel checker.
Demand-brake versus share-shift taxonomy. Pre-define metrics (usage, pricing, win/loss versus OpenAI/Grok/OSS) before the first Anthropic soft print. Expert: product-led AI researcher + channel checks.
Token unit economics / subsidy test. Baker’s central All-In foil is “tokens are subsidized.” Independent view of Anthropic/OpenAI/xAI gross margin per token and capex payback (Sacks’s “as quick as one year” claim). Expert: AI infra sell-side + former hyperscaler finance.
Nvidia residual / AI-factory term sheets. Guaranteed offtake floor, tenor (3–4 years), revenue-share above floor, who bears the “25%” gap risk, which of the six PE/bank platforms are live versus MOU. Data: Jensen essay (10 Aug), CNBC joint interview, Morgan Stanley note Baker cites (not retrieved 24 Aug). Expert: structured-credit / equipment-finance PM.
P2 — needed to underwrite sleeves and debates
Nvidia forward P/E “10-year low.” Confirm as-of first week of Aug 2026 versus history and versus Liberation Day / DeepSeek prints. Press is directionally consistent (BofA 8 Aug note via secondary: ~16× 2027 EPS versus ~36.9× 10-year average — not a primary Bloomberg pull). Expert: equity-linked desk with Bloomberg history.
Physical feasibility of $400–500B ARR. Translate ARR to GW, turbines, interconnects, Texas audit constraints. Does Baker/Sacks zone clear physically while $1T does not? Expert: power markets + DC developer.
OSS versus frontier revenue mix. Test 65–85% value / ~80% volume versus Jason’s “90% cheaper kills premium.” Data: OpenRouter/public pricing, enterprise RFP anecdotes, Meta/GLM adoption. Expert: CIO survey shop.
Grok 4.6/4.7 independent evals. Replicate Databricks-class benches without Cursor conflict; track vibes (DHH etc.) for 2–4 weeks post-4.7. Expert: internal eval harness.
SpaceX sum-of-parts with explicit Grok leg. How do mark-to-market methodologies treat xAI today? What would Anthropic-like ARR scale imply? Expert: late-stage space/AI crossover (conflict-aware — show participants are early SpaceX investors). Confirm MW energized, $/GW, consensus $73B.
Memory LTA terms and break history. What do 2026 LTAs specify (prepay, floor/ceiling, volume)? Any historical breaks and allocation punishment? Hynix/Micron commentary. Expert: memory analyst; supply-chain counsel.
Workday/Silver Lake and software PE wave. Confirm deal terms (~$43B pre-news equity value; talks ongoing, not signed). Whether peers see bids, and whether OSS is actually the thesis PE is underwriting. Expert: software PE partner + public software analyst.
Regulation/energization dashboard. NY moratorium status and copycats; Texas audit milestones; Trump pledge compliance versus residential rates near large DCs; water-usage literature corrected for 10,000× / 100,000× claims. Expert: power markets + local regulatory counsel.
P3 — process, ASR, side-channels
ASR/identity cleanup. Resolve “Yoc’s” agentic-user estimate; “Li Bu” dark horse; Fund AI spelling; Muse 1.1 versus other Meta names; “Fable 5”; DART $6B; “9 figure double Lindy reverse” hyperscaler anecdote; Eric Fischria; Merkor eval. Keep unresolved names out of LP-facing lists.
OpenAI MoM growth hearsay. Sacks’s >20% MoM for 2 months is high-impact if true — triangulate before it enters any model.
China DUV. Context only. Track Shanghai Aishengna deliveries versus the 5 / 20 unit sketch; ASML China order mix; do not upgrade to a 2026–27 earnings thesis without order-book evidence.
Data to collect this month. (i) A three-vintage GPU rental panel: 2024–25 contracted, current reserved, current spot, by SKU (B200, H100, A100). (ii) MSFT/META/AMZN OCF bridge workbook with unusual-item flags. (iii) Anthropic run-rate timeline: $9B YE25 → $47B May → ~$65B Jul-end Reuters → panel $100–120B YE26, with a column for source quality. (iv) State moratorium / audit tracker. (v) Residual-language excerpt file (Huang essay, Nvidia 8-K/press, any MS note).
Expert interviews (who / why). Nvidia IR or a structured-credit PM who has seen an AI-factory term sheet (wrapper existence). A neocloud CRO (contract versus spot). A memory-supply-chain person on LTA break penalties. A power-markets lawyer on EO 62 and ERCOT Batch Zero. An eval engineer with no Cursor affiliation (Grok 4.7). A former software banker on Anthropic lockup/pricing (FT leak as theater versus demand). Do not email anyone from this desk as part of this note.
Primary (transcript / memo).
- Gavin Baker with Patrick O’Shaughnessy, Invest Like the Best EP.485, published 4 Aug 2026, duration 1:18:44. https://www.youtube.com/watch?v=NGsi2PC4y68. Desk memo: /workspace/pm-memos/2026-08-04-gavin-baker-invest-like-the-best-ep485.md. Transcript: /workspace/gavin-baker/transcripts/NGsi2PC4y68.md (ASR; timestamps estimated).
- Gavin Baker with Jason Calacanis and David Sacks, All-In, published 14 Aug 2026, duration 1:39:29. https://www.youtube.com/watch?v=kVzYGVJ8zUk. Desk memo: /workspace/pm-memos/2026-08-14-gavin-baker-all-in-podcast.md. Transcript: /workspace/gavin-baker/transcripts/kVzYGVJ8zUk.md (ASR; chapter-anchored times).
External checks, accessed 24 August 2026.
| Check | Status | URL / locus | Date accessed |
|---|---|---|---|
| 1. MSFT+META+AMZN OCF path (28→32 / adj. 28→35) | Partial. TTM OCF ~$475B and accelerating; FCF compressed. Specific 28→32 vintage and EU-fine add-backs not reconstructed. | Microsoft EX-99.1 https://www.sec.gov/Archives/edgar/data/789019/000119312526323632/msft-ex99_1.htm (OCF $55.4B qtr / $182.9B FY); Meta Q2 https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/ ($31.86B qtr OCF; TTM $130.3B via stockanalysis.com); Amazon EX-99.1 https://www.sec.gov/Archives/edgar/data/1018724/000101872426000024/amzn-20260630xex991.htm (TTM OCF $161.4B, +33%; TTM FCF −$7.6B) | 24 Aug 2026 |
| 2. Nvidia forward P/E vs 10-year; July drawdown | Partial. Press consistent with compressed forward multiple (BofA 8 Aug note via stockwirex: ~16× 2027 EPS vs ~36.9× 10-yr avg; Reuters 30 Mar had 7-year low; StockTi 3 Aug ~22.3×). Not a Bloomberg history audit. SOX July: ~−20.6% month / ~−28.6% peak 22 Jun → trough 29 Jul (quantflowlab, stockwirex, historyofmarket). $1.3T chip-cap wipe in late-Jul sessions (CNBC/FactSet via techstartups). Single-name 40–60% not uniformly confirmed. | https://stockwirex.com/analysis/nvda-nvidia-stock-analysis-decade-low-pe-august-2026/ ; https://www.reuters.com/business/nvidias-pe-sinks-seven-year-low-war-ai-angst-weigh-2026-03-30/ ; https://quantflowlab.com/semiconductor-cycle-history/ ; https://stockwirex.com/analysis/chip-sector-selloff-ai-trade-august-2026/ ; https://techstartups.com/2026/07/29/ai-chip-stocks-lose-1-3-trillion-as-nvidia-tsmc-samsung-hit-by-fears-over-ai-infrastructure-returns/ | 24 Aug 2026 |
| 3. Anthropic IPO / FT $2T / YE26 ARR / October | Partial. FT leak syndicated (Financial Post, QZ, TNW): six backers, October, $100–120B YE26, $2T or more, ~16–20×, confidential S-1 in June, last company figure ~$47B May, May post-money ~$965B. Reuters/Bloomberg 17 Aug: July-end run-rate ~$65B — below Baker’s on-show “>$80B.” $2T is investor expectation, not a company target. October not confirmed by the company (quiet period). | https://financialpost.com/financial-times/anthropic-investors-bet-valuation-ipo ; https://qz.com/anthropic-ipo-2-trillion-valuation-october-081326 ; https://www.reuters.com/technology/anthropic-revenue-run-rate-tops-65-billion-source-says-2026-08-17/ ; https://www.bloomberg.com/news/articles/2026-08-17/anthropic-revenue-run-rate-surpasses-65-billion-ahead-of-ipo | 24 Aug 2026 |
| 4. Nvidia ~$500B AI-factory; Jensen essay; MS residual/royalty | Confirmed on platforms and essay. Partial on residual 25% and MS royalty math (secondary writeups; original MS note not retrieved). | https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital ; https://blogs.nvidia.com/blog/nvidia-ai-factory-compute/ ; https://www.cnbc.com/2026/08/10/nvidia-wall-street-asset-managers-500-billion-ai-push.html ; https://businessmodelanalyst.com/nvidia-usage-linked-revenue-25-percent-backstop/ | 24 Aug 2026 |
| 5. GPU rental/spot vs contract (B200) | Partial. Public list: CoreWeave B200 ~$4.26/GPU-hr spot, ~$8.60 on-demand (computeprices.com 20 Aug; CoreWeave pricing page). Floor ~$2.80 (Genesis). Does not confirm Baker’s contracted mid-$2 vintage or +100% expiry anecdote. | https://www.coreweave.com/pricing ; https://computeprices.com/providers/coreweave/gpus/b200 | 24 Aug 2026 |
| 6. NY moratorium / Trump pledge / Texas audit | Confirmed. NY EO 62, 14 Jul 2026. Texas Abbott letter 3 Aug 2026, 474 GW queue. Pledge expanded ~23 Jul 2026, voluntary, White House 80% coverage claim. | https://www.governor.ny.gov/executive-order/no-62-establishing-temporary-moratorium-data-centers-new-york-while-state-develops ; https://www.texastribune.org/2026/08/03/texas-data-center-project-audit-greg-abbott/ ; https://www.whitehouse.gov/releases/2026/07/president-trumps-ratepayer-protection-pledge-secures-american-ai-dominance-protects-consumers/ ; https://www.reuters.com/legal/litigation/trump-set-expand-power-cost-pledge-data-centers-2026-07-22/ | 24 Aug 2026 |
| 7. Grok 4.6/4.7; Cursor/xAI | Confirmed 4.6 (12 Aug) and Cursor acquisition. Partial on 4.7 (Musk 3–4 weeks from 12 Aug; no official ship date). | https://x.ai/news/grok-4-6 ; https://cursor.com/blog/grok-4-6 ; https://cursor.com/blog/joining-spacex ; https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf | 24 Aug 2026 |
| 8. Workday / Silver Lake; WDAY +17% | Confirmed as Reuters talks (13 Aug) and +17–18% session (close $206.45). Not a signed deal. ~$43B pre-news equity value. | https://siliconangle.com/2026/08/13/workdays-stock-jumps-17-report-silver-lake-buyout-discussions/ ; Reuters syndication via Economic Times 14 Aug | 24 Aug 2026 |
| 9. CoreWeave Ampere / 2029 | Confirmed as company claim (CFO Agrawal on Q2 call ~11 Aug; A100 contract into 2029 “at an attractive price”). One contract, not a fleet proof. | https://www.businessinsider.com/coreweave-deal-challenges-nvidia-ai-chip-obsolescence-fears-2026-8 ; Tom’s Hardware / runtimewire coverage | 24 Aug 2026 |
| 10. SpaceX public listing | Confirmed. Nasdaq SPCX, 12 Jun 2026, ~$135 / ~$1.77T, ~$75B raise. Memo treating SpaceX as public is correct in this timeline. | https://www.space.com/space-exploration/usd1-77-trillion-spacex-is-about-to-become-the-7th-most-valuable-american-company | 24 Aug 2026 |
| 11. China DUV / ASML | Partial as reported production start (Reuters 27 Jul citing The Information): Shanghai Aishengna, ~5 units 2026 / ~20 2027, SMIC/Hua Hong/CXMT; ASML ~8% down on print; China ~16% of ASML H1 sales in press. Unproven in HVM; EUV monopoly intact. | https://www.reuters.com/world/china/china-begins-making-homegrown-duv-chipmaking-tools-information-reports-2026-07-27/ ; https://techwireasia.com/2026/07/china-duv-lithography-asml/ | 24 Aug 2026 |
| 12. Knowledge-work TAM ~$25T | Unverified as TAM. Partial rhyme: McKinsey 2023 $6.1–7.9T gen-AI benefits; later MGI ~$28.7T global automation potential. Not the same construct. | https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier ; https://www.mckinsey.com/mgi/our-research/agents-robots-and-us-skill-partnerships-in-the-age-of-ai | 24 Aug 2026 |
| Bonus: Meta credit scare | Partial. CDS ~95bp (from ~56bp YE25) in late-Jul press; Sopaipilla $12.547B at 7.534% / T+287.5, 1.6× coverage, 27 Jul. | https://finance.yahoo.com/markets/stocks/articles/wall-street-picks-ai-winners-140735436.html ; https://finance.yahoo.com/technology/articles/investors-dig-12-5b-data-134615348.html | 24 Aug 2026 |
| Bonus: Databricks $190B | Confirmed (company, 13 Aug): $5B at $190B, Coatue/BX/MGX/T. Rowe/Sixth Street; >$7B revenue run-rate. | https://www.reuters.com/legal/transactional/databricks-raises-5-billion-financing-190-billion-valuation-2026-08-13/ ; https://www.databricks.com/company/newsroom/press-releases/databricks-grows-80-yoy-surpasses-7b-revenue-run-rate-scales | 24 Aug 2026 |
Reconstructed hyperscaler cash (desk, 24 Aug 2026; not Baker’s 28→32).
| Name | OCF, qtr ended 30 Jun 2026 | OCF, TTM/FY | FCF tell |
|---|---|---|---|
| Microsoft | $55.4B (vs $42.6B) | $182.9B FY (vs $136.2B) | PP&E additions $35.8B qtr / $115.9B FY |
| Meta | $31.9B (vs $25.6B) | $130.3B TTM | FCF $0.8B qtr; PP&E $30.1B |
| Amazon | n/a as a clean qtr in the release extract | $161.4B TTM (+33%) | TTM FCF −$7.6B; PP&E net TTM $169.0B |
| Sum (TTM-ish) | — | ~$475B | FCF compressed / negative |
July tape versus Baker telemetry (compressed).
| Object | Tape / press | Baker private | Desk |
|---|---|---|---|
| SOX | ~−21% July / ~−29% peak-to-trough | AI names −40–60% | Index confirmed; 40–60% single-name unverified as a uniform fact |
| Demand | ROI/financing fear | GPU, DRAM, tokens accelerated | Telemetry unaudited |
| Meta | Rent compute = glut | No CapEx cut; telemetry hotter | Both can be true at different horizons |
| OSS | Mix = over | Mix = infra margin transfer | Measure dollars vs tokens |
| Credit | CDS/bonds ugly | Only real July fundamental | Partial confirm on Meta CDS / Sopaipilla |
| China DUV | Semi-cap selloff | Phase change, 25 years behind, years to order book | Reported 5/20 units; do not overclaim |
Baker ≠ Calacanis ≠ Sacks. Pie growth, reprice bridge, OSS-as-infra, S1 as subsidy-falsifier, Grok-in-SpaceX, regulation-as-#1-risk, LTA Game of Thrones, and the $400–500B YE27 zone are Baker (the last jointly with Sacks). OSS-price-crush, GLM 90% cheaper, slower triple, OpenAI capex $1.4T→$600B, and Bending Spoons color are Calacanis. Fed-of-AI, 6–8 GW / $300–400B, >20% OpenAI MoM, 90-day Anthropic cancel, and FAA-for-AI irony are Sacks. Amazon DSP is not a Baker thesis. Scenario weights are desk inference. No trade recommendations. No emails. No publication.
Prepared 24 August 2026 for internal institutional-research use (AP). Educational synthesis of public interviews plus desk external checks. Not investment advice. ASR and estimated timestamps apply to the primary sources. Speaker-stated numbers are attributed, not independently verified, except where the Appendix marks confirmed or partial.
Thematic research · not a trade recommendation · Mary desk copy