In January 2025, OpenAI, Oracle, and SoftBank stood at the White House and announced Stargate — $500 billion, 10 gigawatts, the largest AI infrastructure commitment in history. The flagship Abilene campus is genuinely operational, running real chips on real workloads. But the three-way joint venture meant to deliver the other nine sites has, by independent reporting, stalled completely — no shared staff, no shared data centers, the partners pursuing separate bilateral deals instead. Actual committed equity to date: $52 billion. The remaining $448 billion was always meant to come from debt not yet raised. And OpenAI's own spending target has already been cut once, from $1.4 trillion to $600 billion, after investor pushback. This is the letter on the difference between an announcement and an asset.
Not investment advice. Data sourced from OpenAI and SoftBank official disclosures, MLQ News February 2026, IntuitionLabs Infrastructure Analysis April 2026, CNBC February 2026, R&D World January 2026, Carnegie Investment Counsel, Senator Elizabeth Warren correspondence January 2026. All figures current as of June 2026.
In January 2025, at a White House event with President Trump, OpenAI, Oracle, and SoftBank announced Stargate: a joint venture to invest $500 billion over four years building 10 gigawatts of AI data centre capacity in the United States — at the time, the single largest privately funded infrastructure commitment in history, larger than the GDP of most countries. Eighteen months later, two genuinely contradictory things are simultaneously true, and understanding both is the entire purpose of this letter.
The first truth: the flagship Abilene, Texas campus is real, operational, and genuinely impressive. Oracle began delivering NVIDIA GB200 racks in mid-2025; OpenAI has confirmed it is running early training and inference workloads on the new capacity; the campus employed more than 6,400 construction workers at peak; the first two of ten planned buildings have been operational since September 2025, with Larry Ellison personally confirming the remaining buildings under active construction. This is not vapourware. Real chips, real power draw, real frontier-model training is happening in West Texas right now.
The second truth, reported independently in February 2026 and not seriously disputed since: the three-way joint venture structure announced at the White House — the actual "Stargate" entity meant to jointly own and operate the remaining nine sites — has stalled completely. No joint venture staff have been hired. No data centres have been developed under the joint venture structure itself. Disagreements over site control and operational responsibility between OpenAI, Oracle, and SoftBank proved unresolvable, and OpenAI instead pursued separate bilateral agreements with each partner individually — a 4.5 gigawatt deal directly with Oracle, a separate compromise with SoftBank reached after negotiations in Tokyo, each bypassing the original joint Stargate framework entirely. The brand survived. The structure announced at the White House did not.
The single most important number in this entire story, and the one that gets least attention relative to the much larger headline figure, is the actual committed equity behind Stargate as reported by The Information: $19 billion each from SoftBank and OpenAI, $7 billion each from Oracle and MGX (the Abu Dhabi state investment vehicle) — $52 billion in total, against a $500 billion headline commitment. The remaining $448 billion was never intended to come from these four partners' own balance sheets. It was always structured to come from debt — project financing secured against the data centres themselves, the same model used to finance large-scale infrastructure like power plants and toll roads — and that debt has not yet been raised at anything close to the scale required.
SoftBank's own CFO conceded in August 2025 that the project was "taking longer than anticipated" — a notably understated characterisation from one of the joint venture's four equity partners. Oracle's credit profile has deteriorated under the weight of its Stargate-related capital commitments specifically; Barclays downgraded Oracle's stock to "underweight" in early 2026 and explicitly warned the company's credit rating could approach junk status if infrastructure spending continues outpacing revenue conversion from the underlying AI workloads. This is not a peripheral observation — Oracle is one of Stargate's four core partners, and a credit downgrade at one of the entities meant to help finance $448 billion in unsecured debt materially worsens the prospects for raising that debt at favourable terms.
The gap between the press releases and the underlying corporate reality
A structural pattern running underneath Stargate deserves its own examination, because it complicates any simple reading of the headline investment figures. Microsoft has invested over $13 billion in OpenAI and continues providing cloud services to it; OpenAI commits to buying Microsoft cloud capacity in return. Nvidia has committed up to $100 billion in OpenAI — money that, as OpenAI's own CFO Sarah Friar has publicly acknowledged, "will go back to Nvidia" in the form of GPU purchases. Nvidia is also a prominent investor in CoreWeave, which itself supplies cloud capacity to OpenAI and has spent billions of its own buying Nvidia chips.
The market reaction to this concentration risk has already happened once, and it was severe. On January 29, 2026, Microsoft's stock fell 12% in a single session — erasing $440 billion in market value — specifically after the company disclosed that 45% of its entire $625 billion cloud services backlog was tied to OpenAI alone. This is the practical, market-tested consequence of circular AI financing: when one company's commitments make up nearly half of a trillion-dollar backlog at a separate, much larger public company, that company's own equity becomes a leveraged bet on the smaller company's survival — a risk that does not show up cleanly in either company's own headline financial statements, but that the market priced in immediately and brutally once it understood the concentration.
The implied growth rate required for OpenAI's $280 billion 2030 revenue target — from roughly $13 billion today — is, by independent analyst calculation, equivalent to sustaining 100% year-on-year revenue growth for the entire remainder of this decade. The closest historical precedent most analysts cite is Google's early-2000s search advertising growth — a company that grew from near-zero to internet-defining scale by capturing advertising dollars away from an entire legacy industry (print, radio, broadcast television) almost in its entirety. Whether OpenAI's consumer chatbot and enterprise API business has a comparably large, comparably capturable revenue pool to displace remains genuinely unresolved — and unlike Google's advertising moat, OpenAI faces direct, well-funded, and increasingly cost-competitive rivals in a way Google did not in 2003.
Similarweb's Global AI Tracker data shows ChatGPT's web traffic share falling from 86.7% in January 2025 to 64.5% in January 2026 — a 22-percentage-point decline in a single year, even as the overall AI chatbot market itself expanded substantially. DeepSeek V3.2, released December 2025, matches GPT-5-level performance on elite reasoning benchmarks at inference costs reported to run 10 to 30 times cheaper. Kimi K2.5, a trillion-parameter open-source model from Moonshot AI, reportedly beats GPT-5.2 on tool-augmented reasoning tasks at roughly one-twentieth the cost. The competitive and cost pressure on OpenAI's core product is intensifying at precisely the moment its infrastructure spending commitments are at their largest in the company's history.
OpenAI carries the largest, least diversifiable exposure of any Stargate partner — it is simultaneously the joint venture's primary customer, a direct equity investor, and the entity whose revenue growth must justify the entire infrastructure buildout. The company's March 2026 funding round, raising $122 billion at an $852 billion post-money valuation with participation from Amazon ($50B), Nvidia ($30B), and SoftBank ($30B), pushed total equity raised to date past $170 billion — narrowing but not closing the gap against the original $1.4 trillion five-year compute cost projection, even after that figure was cut to $600 billion.
Senator Elizabeth Warren's January 2026 letter to Sam Altman, requesting details on OpenAI's spending commitments and the circumstances under which the company might seek federal "backstop" support, captures the structural concern precisely: a private company with negative free cash flow, sustained 57% cash burn rates, and trillion-dollar-scale infrastructure commitments is, by the company's own CFO's acknowledgement of a potential government backstop, raising the question of who bears the loss if the revenue projections do not materialise.
Oracle has positioned itself as the primary physical builder and cloud operator behind Stargate's flagship Abilene campus and several subsequent sites, with its order backlog reportedly swelling to $523 billion on the strength of its AI infrastructure commitments. This is simultaneously Oracle's largest growth opportunity in a generation and its most acute single-customer concentration risk — the same dynamic that triggered Microsoft's 12% single-day stock decline applies with comparable force to Oracle, whose own credit profile Barclays has already flagged as deteriorating specifically because of Stargate-related capital intensity outpacing confirmed revenue.
SoftBank's role combines direct equity investment with site development expertise — repurposing a former Foxconn-associated auto plant in Lordstown, Ohio into what the consortium describes as a proving ground for advanced modular data centre design, alongside a separate 1.2 gigawatt site in Milam County, Texas developed through its SB Energy subsidiary. SoftBank's own CFO publicly conceding the project was "taking longer than anticipated" is among the most candid statements any Stargate partner has made on the record — notable from a company with Masayoshi Son's history of aggressive, occasionally overextended technology infrastructure bets (WeWork being the most cited cautionary precedent).
The three companies most directly implicated in the circular financing dynamic examined above — each simultaneously an OpenAI investor, an OpenAI supplier, and in Microsoft's case, an OpenAI cloud services customer-of-its-own-customer. Microsoft's January 2026 stock decline is the single clearest piece of market evidence available that investors are beginning to price circular AI financing as genuine concentration risk, not merely a footnote to otherwise strong AI-driven earnings growth.
Stargate's flagship site being genuinely operational does not validate the headline $500 billion figure, and Stargate's joint venture structure stalling does not mean OpenAI's compute buildout is failing. Both extremes — "Stargate proves the AI infrastructure boom is real and unstoppable" and "Stargate proves the entire AI infrastructure narrative is a bubble" — collapse two genuinely separate questions into one. Whether Abilene works as a data centre is an engineering and operations question, already substantially answered: yes, it works. Whether the broader $500 billion, 10-gigawatt commitment gets built as originally announced, through the originally announced joint venture structure, on the originally announced timeline, is a financing and corporate-governance question — and the honest answer, based on February 2026 reporting that has not been seriously disputed, is no, not as originally structured.
The reflexive comparison to Google's early-2000s revenue growth, which OpenAI's bulls cite to justify its 2030 targets, understates one critical difference: Google operated in a market with essentially no well-funded, technically comparable rival during its fastest growth years, while OpenAI faces DeepSeek, Kimi K2.5, Anthropic, Google's own Gemini, and Meta's Llama simultaneously, several offering reasoning-benchmark-comparable performance at a fraction of the inference cost. A 100% sustained annual growth rate is achievable for a company with a genuine moat against an unfunded competitive field. It is a considerably harder bet for a company whose own former chief scientist (Ilya Sutskever) has publicly stated that the pre-training scaling approach underlying its core product may be approaching diminishing returns, and whose flagship consumer product's web traffic share has already fallen more than 20 percentage points in a single year.
The circular financing structure examined in this letter is not inherently fraudulent or even unusual for a capital-intensive emerging industry — railroads, telecommunications, and the original dot-com infrastructure buildout all featured comparable supplier-investor-customer overlaps — but it does mean that the headline investment figures circulating in AI infrastructure coverage systematically overstate genuinely independent, arm's-length capital commitment. When Nvidia's $100 billion OpenAI investment is substantially expected to flow back to Nvidia as chip revenue, and Microsoft's cloud backlog growth is 45% dependent on a single customer it also helps finance, the appropriate analytical response is not panic, but the same discipline this letter series applied to AI hyperscaler bond issuance in Letter 110: distinguish between capital that has genuinely changed hands at arm's length and capital that is, in substantial part, the same dollars moving between a small number of deeply intertwined balance sheets.
Long-horizon thinking on capital, technology, and the forces shaping the next decade of wealth creation. Written from first principles. Not consensus. Not noise.