The Investor

The Structure

For the first time this week the full circular financing geometry of frontier AI is actually measurable, which is either useful or depressing depending on your book. Anthropic committed $200B to Google Cloud over five years, more than 40% of GCP's disclosed backlog, and Google simultaneously put up to $40B in equity back into Anthropic. OpenAI, not to be outdone on the geometry of the loop, committed $688B across Microsoft ($250B), Oracle ($300B), and Amazon ($138B). Those same four hyperscalers have separately written $88B+ in equity into the two labs.

Combined cloud commitments run to $1.018 trillion, roughly half of the $2T+ cloud backlog disclosed across Microsoft, Oracle, Google, and Amazon. The money the sellers handed to the buyers is now booked as the sellers' revenue growth.

One analyst's framing captures it precisely: 'This is either the largest working-capital loop in the history of enterprise software or an ordinary vendor-finance arrangement in louder clothing. It is probably closer to the second, which is not as reassuring as it sounds.'

Why the Surface Read Is Insufficient

The consensus bull case says revenue growth validates the spend. It has two empirical problems. Anthropic's $330B commitment equals ~10x its current run-rate revenue, and OpenAI's $688B ratio is worse. Then there is a February 2026 paper (Kapoor et al.) that tested 14 frontier models over 18 months and found capability improved substantially while reliability barely moved. Power users consume 7x more than median users, and those same power users churn providers every two months, which means the run-rate figures are probably double-counting the highest-value cohort across both labs.

Dario Amodei's defense: 'The other player is pretty confident they'll have the revenue at the right time.' That sentence is load-bearing for roughly a trillion dollars of commitments.

Counterparty Fragility Matrix

The Three Scenarios Worth Pricing

Scenario 1 (consensus): revenue absorbs the capacity, nobody cares about circularity, infra names grind higher. Scenario 2: hyperscalers miss cloud revenue growth, the loop gets redescribed as something less flattering, and apps outperform by default. Scenario 3 (most likely, or rather the more interesting version): the loop keeps spinning but markets begin pricing layers differently because someone, eventually, has to show cash conversion. Sources disagree on which scenario dominates and converge on one point, which is that the layer insulated from the loop unwind is where the alpha sits.

That layer, in practice: reliability and eval infrastructure, vertical data moats in regulated workflows, and the physical enablers (power, cooling, advanced packaging) that get paid regardless of which model wins.

The Setup

Cerebras prices Tuesday, May 13, running limit-order IPO mechanics to manage a book that is plainly too hot for a conventional allocation process. A $3.5B raise at a $26.6B valuation, with $10B in orders, which works out to a 2.86x book. Largest tech IPO of 2026 and the first clean public print for inference-optimized silicon since the 2024 freeze.

The cap table is where it gets interesting. Sam Altman, Greg Brockman, Ilya Sutskever, and Intel CEO Lip-Bu Tan wrote personal checks. The Intel CEO angel-investing in a Nvidia challenger is not a press release. It is someone with an enormous day job telling you what he thinks his day job is worth in five years.

The OpenAI Concentration Problem

OpenAI sits on the other side of a $10B+ multi-year contract, a $1B loan, and a 33M share purchase option. A non-trivial slice of the enterprise value being marketed here is, functionally, a claim on the continued health of a single private company currently in litigation with Elon Musk and losing senior talent at a rate the S-1 does not quantify.

Call this vertical integration by financial instrument rather than by M&A. It is probably the template for the next cycle, or rather, the version of it that gets written into the next set of term sheets. Anthropic will likely mirror it within two quarters with its own anchor customer attaching the same mix of contract, debt, and equity optionality. That reshapes how AI infra deals price, and — this is probably wrong, but — it also reshapes which buyers can credibly participate.

If Cerebras breaks issue cleanly, the calendar behind it — SPX, OpenAI, Anthropic, and late-stage infra names sitting on stale marks — gets pulled forward by a quarter or two, and secondaries reprice 20-40% within 30 days. If it breaks issue messily, the window closes again.

What This Means for the Private Book

The $26.6B print becomes the anchor comp for Groq, SambaNova, Tenstorrent, Rain AI, and for every other inference-silicon position sitting in LP books at stale 2023 marks. The market is also starting to price inference silicon as a different animal from training silicon, which means the thesis memos that have been conflating the two need to stop doing that.

The near-term trades sort themselves by how quickly the edge decays.

1. Pre-IPO secondaries (48-hour window). Anyone with LP relationships on adjacent cap tables is already moving before pricing. The comp repositions those marks 20-40% inside 30 days if Cerebras trades well, and the sellers know it as clearly as the buyers do.
2. The customer-as-investor-as-creditor template (this quarter). The infra startups worth a second look are the ones where a hyperscaler has signed a multi-year contract but has not yet attached equity or debt. Those tend to get Cerebras-style structured capital inside twelve months, or the anchor customer finds a reason not to renew.
3. Inference ≠ training as a permanent thesis bifurcation. LPs will ask about this on the next annual call. The GPs with a crisp answer set the narrative. The ones still quoting H100 comps do not.

Commoditization From Above: Apple Opens the Marketplace

Apple's iOS 27 ships this fall and will let users pick from multiple third-party AI models for text and image work, which is precisely the opposite of the exclusive-partner posture Cupertino telegraphed at WWDC 2024. For the foundation labs this is free distribution to over a billion devices. For consumer AI wrappers, whose entire pitch was a distribution wedge Apple would never offer natively, it is closer to a funeral notice.

The arithmetic underneath is the part worth staring at. Google's Gemini pulled 22M+ downloads against ChatGPT's 12M on image launches, and ChatGPT still booked $70M in revenue while the rest booked something rounding to nothing. Distribution is not converting to economics for anyone except OpenAI.

Commoditization From Below: Open Source Hits Parity

An independent benchmark, DeepResearch Bench, ran 100 PhD-level tasks and put open-source stack Onyx at #1, above OpenAI, Gemini 2.5 Pro, and Perplexity. Separately, Subquadratic claims 12M-token context at roughly 5% of Opus pricing with a 50M-token follow-up teased. If even half of that survives replication, long-context pricing, the tier where labs charge most and discount least, stops being a moat.

Reasonable people disagree on whether SubQ's claims hold up under scrutiny. The consensus, or rather the more honest version of it, is that even if SubQ disappoints, the direction of travel is clear. RAG infrastructure and generic agent tooling are being disintermediated regardless.

The Labs' Escape Route: Services

Both frontier labs launched PE-backed services JVs this week. Anthropic's is $1.5B with Goldman/Blackstone/H&F. OpenAI's is a $4B 'Deployment Company' at $10B pre-money with 19 investors, and Brad Lightcap has moved from COO to run it. The structure is a quiet admission that API revenue alone cannot hit their growth targets. The next $10B of enterprise AI value accrues at the deployment layer, which is a polite way of saying consulting.

Anthropic disclosed finance as its #2 revenue segment, shipped 10 production financial agents, and named Goldman, Visa, Citi, and AIG as reference customers. OpenAI took the other fork entirely: $100M ad ARR in 6 weeks, CPC self-serve ads inside ChatGPT, a stated target of $2.5B for 2026.

Where Value Migrates

The thesis across 10+ sources has been forming for a while and is unchanged this week: model-layer margins compress, orchestration and vertical-workflow margins expand. This is probably wrong in places. Where the money actually accrues over the next eighteen months is not mysterious.

1. Vertical agents outside BFSI. Anthropic validated the template in finance. Insurance, legal, healthcare, and industrial sit uncontested for 12-18 months before someone else gets serious, which is longer than most funds assume.
2. Data infrastructure picks-and-shovels. Only 15% of enterprises are data-ready for agentic AI despite spending tens of millions preparing to be, and 73% name data connectivity as the #1 blocker. That gap is a business.
3. Grounding and hallucination infrastructure. A 30% ungrounded-claim rate on best-in-class models is a structural ceiling, not a bug that scales away. Something has to sit on top of it.