이미지: The Decoder
Summary
- NVIDIA has cut its guarantee for OpenAI's Ohio data center from $25 billion to under $12 billion
- The guarantee covers only the roughly 5-gigawatt Phase 1 build, while the full 10-gigawatt project is being negotiated separately
- Anthropic's revenue rose from $4.73 billion in Q1 to $11.5 billion in Q2, a 14x increase year over year
- 엔비디아 보증액 변화
- 2500억달러 → 1200억달러 미만
- 보증 대상 용량
- 1단계 건설, 약 5기가와트
- 전체 프로젝트 규모
- 10기가와트, SB에너지(소프트뱅크 자회사) 개발
- 별도 칩 구매 금융 협상
- 최대 3500억달러
- 앤트로픽 분기 매출
- 1분기 47억3000만달러 → 2분기 115억달러(전년比 14배)
- 앤트로픽 2028년 매출 전망
- 1900억~2000억달러(5월 공개 연환산 450억달러 대비 대폭 상향)
- 앤트로픽 IPO 계획
- 9월 말~10월 초, 밸류에이션 약 1조달러(알려짐)
NVIDIA cuts its guarantee in half
The scale of the guarantee NVIDIA had pledged for the large data center OpenAI plans to build in Ohio has been significantly reduced. According to a Wall Street Journal report, NVIDIA scaled back its original plan to guarantee $250 billion to less than $120 billion — a cut of nearly half. Investor pressure was behind the move. Concerns arose that NVIDIA was taking on too much risk tied to a single customer, and the company responded by reducing its exposure.
The reduced guarantee applies not to the entire data center but to the Phase 1 construction, which covers roughly 5 gigawatts of capacity. The full project spans 10 gigawatts and is being developed by SB Energy, a SoftBank subsidiary. OpenAI is negotiating a lease covering that full capacity separately from NVIDIA's guarantee. In addition, NVIDIA is reportedly discussing separate financing of up to $350 billion to support OpenAI's chip purchases. Rather than one single mega-deal, the arrangement consists of guarantees, leases, and financing running on separate tracks simultaneously.
NVIDIA makes the chips that power AI, and rather than competing directly in the chatbot race with products like ChatGPT or Claude, it sells the computing infrastructure that race depends on. OpenAI, for its part, develops its own GPT-series models but relies on Microsoft, NVIDIA, and others to lease the data centers and chips needed to run them. The reduced guarantee means NVIDIA has taken a step back from some of the risk in that "leasing" relationship.
Anthropic's revenue jumps 2.4x in a single quarter
Numbers released by Anthropic around the same time point in the opposite direction. According to Reuters, Anthropic's revenue jumped from $4.73 billion in Q1 to $11.5 billion in Q2 — a 14x increase from the same period a year earlier. According to people familiar with the company, Anthropic is projecting 2028 revenue in the range of $190–200 billion, far above the roughly $45 billion annualized revenue figure it disclosed last May. Anthropic said it has grown revenue more than tenfold every year for three consecutive years through early 2026.
Anthropic is reportedly planning to go public at a valuation of around $1 trillion as early as late September or early October. However, payments company Ramp, which measures corporate customers' Anthropic token usage, found signs that demand growth has somewhat moderated.
Signals from the two companies, in a table
| Category | NVIDIA·OpenAI | Anthropic |
|---|---|---|
| Direction | Reduced risk exposure | Surging revenue |
| Key figure | Guarantee cut from $250B to under $120B | Quarterly revenue $4.73B → $11.5B |
| Background | Investor pressure | 14x year-over-year growth |
| Market reading | Reinforces bubble concerns | Counters bubble narrative |
As METAL LAB reported on August 14, OpenAI and Anthropic are also engaged in aggressive price cuts as they face pressure from Chinese AI models catching up. GPT-5.6 Luna cut both input and output token prices by 80% each, and Claude Opus 5 launched at half the price of the top-tier Fable 5 model. That Anthropic's revenue surged this much even under such pricing pressure suggests that, at least on the demand side, the market remains hot.
Editor's take
Placed side by side, these two stories from the same week raise an interesting point. NVIDIA didn't cut its guarantee because it doubts OpenAI. Investors asked whether it made sense to bet the company's entire risk on a single customer's revenue, and NVIDIA simply answered that question. Anthropic's numbers, meanwhile, show that demand itself hasn't died down. The two stories may seem contradictory, but they actually operate on different levels. One is a question of financial structure — who bears how much risk — while the other is a question of demand — whether the number of paying customers is actually growing.
This kind of adjustment isn't new. A recurring pattern has emerged where numbers get trimmed a few months after major infrastructure announcements. The initial announcement presents the maximum figure, and once contracts are actually structured, the numbers get readjusted to spread out risk. This case appears to follow that same pattern. What makes it a bigger story this time is that the cut was nearly 50%, and that it happened to coincide with Anthropic's announcement of explosive growth.
There are two points worth noting for domestic companies. First, the headline totals announced for major AI infrastructure deals shouldn't be taken at face value — it's worth separately tracking how much gets trimmed once deals are actually executed. Second, it's important to distinguish between companies like Anthropic, where demand itself is surging, and companies like NVIDIA, which are spreading out risk — these send entirely different signals. Trying to sum up the bubble debate in a single sentence is itself misguided.
In the coming weeks, the lease terms for OpenAI's full 10-gigawatt project and the outcome of NVIDIA's $350 billion chip financing negotiations are likely to be revealed one after another. If either is reduced or delayed, the bubble debate will intensify again; if Anthropic's projected 2028 revenue is confirmed by actual results, that adds one more piece of evidence to the opposite side.
