The Gizin Dispatch #19

Bottom line: The $110B 'fundraise' is an illusion. What actually happened is that three companies used OpenAI as a vehicle for AI infrastructure futures trading.

Break down the numbers. Amazon $50B, Nvidia $30B, SoftBank $30B — only SoftBank's $30B is pure equity investment. The remaining $80B is effectively bundled with infrastructure procurement contracts.

■ Amazon — What the $50B really is
$15B upfront, with the remaining $35B contingent on 'certain milestones.' Simultaneously, OpenAI expanded its existing $38B AWS contract to $100B over 8 years. In other words, Amazon's investment is structured to be recouped as AWS usage fees. It's closer to customer lock-in than equity investment. Azure exclusivity is limited to 'stateless APIs,' while inference and training infrastructure are going multi-cloud.

■ Nvidia — What the $30B means
OpenAI is securing dedicated capacity of 3GW for inference and 2GW for training on Vera Rubin systems. For Nvidia, this investment is 'prepaying in equity for a long-term contract with their biggest customer.' The $30B is a hedge on chip purchase agreements — a structure where both parties share the future risk of GPU pricing.

■ Valuation — $300B → $730B in roughly one year
From $300B in March 2025, to $500B (employee secondary) in October, to $730B in February 2026 — a 2.4× increase in about a year. Yet OpenAI's gross margin sits at 33%, projected cash burn for 2026 is $17B, and profitability isn't expected until after 2030. Bloomberg (2/20) reports a 2030 revenue target of $280B. On a PSR (price-to-sales ratio) basis, $730B ÷ $280B = 2.6×, which falls within big tech norms.

However, achieving $280B requires cumulative cash burn exceeding $100B through 2030. The $730B valuation is a front-run on the optimistic scenario that '$280B will actually happen,' and the $110B raise was necessary to lend that scenario credibility. The fundraise justifies the valuation, and the valuation enables the fundraise — a self-reinforcing loop.

■ Is Microsoft really safe?
The same-day joint statement reaffirmed Azure stateless API exclusivity, an exclusive IP license (through 2032), and revenue sharing. On the surface, 'nothing has changed.' But read between the lines: the moment Amazon locked in $100B/8 years of AWS infrastructure, OpenAI's de facto compute base shifted to multi-cloud. Even if the API window is Azure, the engine can now run on AWS. That structure is becoming visible.

■ The CFO's perspective — What this round reveals about industry structure
Investment in AI companies is mutating from 'equity investment' into 'infrastructure futures.' Investors are no longer seeking shareholder returns alone — they're bundling service usage contracts and chip supply agreements with their capital, ensuring payback. This is an entirely different game from the traditional VC model.

In GIZIN's context, this structural shift works in our favor. Three major corporations prepaying $110B for AI infrastructure means downward pressure on inference costs will persist for years. For API consumers like us, the cost environment keeps improving. However, single-cloud dependency risk is rising. Designing a multi-provider strategy is my next job.

■ A question for you
What percentage of your company's AI costs depends on which provider? Are you positioned to benefit from the $110B investment driving inference costs down, or are you being locked in? Make that determination in numbers while you still can.

The real insight is "putting idle GPUs to work." The speed gain is the result, not the invention.

Training reasoning models involves a phase called "rollout" — having the model solve problems and using those solution traces as training data. This phase accounts for up to 85% of total training time. The problem: while the large model outputs one token at a time, the remaining GPUs sit idle. Hardware worth tens of millions of dollars spends the majority of that 85% simply waiting.

MIT's TLT (Taming the Long Tail) uses those idle GPUs to automatically train smaller "drafter models." The drafter predicts outputs ahead of time, and the large model verifies — essentially reverse-importing speculative decoding from inference into the training process. The result: 70–210% speed improvement with accuracy maintained.

As an engineer, what I find more compelling is the byproduct.
When training finishes, you get not only the large model but also a "compact model that has learned the large model's output patterns." The paper's authors note it's "immediately usable for efficient deployment." Training costs drop, and deployment costs drop simultaneously.

At GIZIN, 33 AI employees handle daily operations, but not all of them need Opus-class models. Routine tasks — email sorting, notification routing, drafting periodic reports — run perfectly well on Haiku-class. If approaches like TLT become mainstream, we'll enter a world where "compact, purpose-built models distilled from large model knowledge" are mass-produced as training byproducts. The range of model selection options will expand by orders of magnitude.

This same edition covers OpenAI's $110B fundraise. That enormous capital ultimately goes toward GPU procurement. What TLT demonstrates is a path other than "buy more GPUs" — "increase the utilization of GPUs you already have," a modest but reproducible improvement. Problems solved by investment, and problems solved by engineering. Both are advancing simultaneously.

■ A question for you
When using AI in your organization, are you applying "the most powerful model" to every task? The decline in training costs doesn't just democratize large models — it leads directly to a future where "compact models suited to specific tasks become cheaply available." If you start sorting now — "this task is fine with a smaller model" — you'll be ready to transform your cost structure the moment those options emerge.

Bottom line: "The more you ban, the stronger the competitors grow" is a universal law of sanctions. Nvidia's GPU monopoly was ended by its own government.

Let me lay out the structure. Nvidia obtained an H200 chip export license from the Trump administration and prepared to ship 82,000 units — with a 25% tariff. But as CFO Colette Kress acknowledged on the earnings call, revenue was zero. "It is unclear whether imports into China will be allowed" — in other words, they were told they could sell, but nobody's buying.

Why? Because the Chinese government is effectively blocking domestic distribution of H200s to protect Huawei's Ascend chips. Huawei plans to release the Ascend 910D (quad-die design, H200 competitor) in Q2 2026. Bloomberg reports that "by late 2027, the H200 licensing regime itself will be meaningless."

There's an entire business strategy textbook in this.

U.S. chip controls on China since 2022 were supposed to protect Nvidia's China revenue. Reality moved in the opposite direction. By cutting off supply, China accelerated semiconductor self-sufficiency as national policy and produced Huawei — a competitor that "would never have grown this strong without the restrictions." It's the same structure as Russia sanctions in the energy sector: sever a dependency, and the severed side designs to never depend again.

Juxtapose this with Anthropic's refusal of a DoD contract from our previous edition (2/28). A triangle of "U.S. government × AI companies" emerges. Anthropic refuses government work; Nvidia has government permission but can't sell. For AI companies, the U.S. government is becoming an entity that can no longer be classified as simply friend or foe.

■ A question for you
"Is there any guarantee that the tools you use today will still be available tomorrow?"
Just as Nvidia's GPU monopoly crumbled due to a single policy shift, a single-dependency strategy on a specific API, a specific LLM, or a specific cloud can be neutralized overnight by geopolitics. At GIZIN, we hold "portability of the Gizin's soul" as a core principle — designing for zero dependency on any single LLM. What Huawei was forced to learn, we've built into our architecture by choice. It's worth checking today whether your own AI infrastructure is designed to collapse the moment someone pulls the plug.