NVIDIA chief executive Jensen Huang told the company’s GTC conference on 16 March that he expects purchase orders for its Blackwell and Vera Rubin systems to reach $1tn through 2027, CNBC reported, double the $500bn opportunity the company had projected for the two chip generations. NVIDIA’s own GTC blog put it slightly differently: Huang “now sees at least $1 trillion in revenue from 2025 through 2027”, with computing demand “off the charts”.
NVIDIA shares rose about 2% on the day, CNBC said. The more useful question for anyone following the build-out is who is paying for that trillion, and the answer is in the spending plans the biggest buyers set out a few weeks earlier. We covered the keynote itself in our GTC preview, our first-day write-up and a longer look at the announcements.
What the biggest buyers plan to spend in 2026
Three of NVIDIA’s largest customers gave 2026 capital expenditure forecasts with their year-end results:
- Amazon: “about $200 billion”, chief executive Andy Jassy said in the results statement, up from roughly $131bn in 2025, according to CNBC. Amazon’s shares fell 8% the next day.
- Alphabet: $175bn to $185bn, against $91.4bn in 2025, the company said on its fourth-quarter earnings call. Alphabet noted that its compute options include “GPUs from our partner NVIDIA”.
- Meta: $115bn to $135bn, including finance-lease payments, against $72.22bn in 2025, according to its fourth-quarter results.
Microsoft did not give a specific forecast for the year, CNBC noted. Taken with Microsoft, the big cloud and platform companies plan to invest a combined $700bn or so in AI this year, CNBC reported separately.
Meta: the steepest rise, and questions about cost
Meta’s plan is the steepest proportionally: at the top of its range, it would almost double 2025’s spending. Days before GTC, Reuters reported that Meta was planning to lay off more than 20% of its staff to balance that spending. A Meta spokesperson told CNBC: “This is a speculative report about theoretical approaches.” Our Meta spending piece covered the report at the time.
Meta is also buying capacity rather than only building it. On the day of the keynote, Nebius disclosed a deal with Meta worth up to about $27bn over five years, in a filing with the SEC: $12bn of dedicated GPU capacity from early 2027, plus up to $15bn for capacity in other clusters that Nebius fails to sell to other customers. Nebius said the dedicated capacity would be one of the first large-scale deployments of Vera Rubin. We looked at that deal in our Nebius story.
What NVIDIA is promising buyers
NVIDIA’s pitch to these customers is cost per unit of output. In its Vera Rubin announcement, the company said the platform’s seven chips are “now in full production”. It says the Vera Rubin NVL72 rack delivers “up to 10x higher inference throughput per watt at one-tenth the cost per token” compared with Blackwell. Those are NVIDIA’s own figures. NVIDIA said partner products would be available from the second half of 2026, naming Amazon Web Services, Google Cloud, Microsoft Azure and Oracle among the clouds, and Nebius among the smaller AI cloud providers.
Why the capex plans favour the supplier, for now
Set side by side, the numbers explain the confidence on stage. NVIDIA sells to all of the companies above, so their capital budgets are its order book. Amazon, Alphabet and Meta between them plan to spend roughly $490bn to $520bn this year on TSN’s arithmetic, before Microsoft, and every one of those plans rose sharply on 2025. A supplier does not need to choose a winner among its customers when all of them are raising spending at once.
The risk sits mainly with the buyers. Amazon’s share price fell after it set out its plan, and Meta faced a report, which it called speculative, that it was planning job cuts to balance its own. If the returns on that spending are slow to arrive, it is the customers’ budgets that come under pressure first, and NVIDIA’s orders would follow.
The cost-per-token pitch cuts both ways. If Vera Rubin delivers what NVIDIA claims, a buyer gets more output for the same money, which supports the spending case. It also means each dollar of capex buys more capacity, so demand has to keep growing for the orders to hold. NVIDIA’s $1tn figure assumes it will.
What has changed since March
Meta has since raised its range. In its first-quarter results it put 2026 capital expenditure at $125bn to $145bn, citing “higher component pricing” and, to a lesser extent, data centre costs for future capacity. Its headcount was 77,986 at the end of March, against 78,865 at the end of 2025. Amazon now plans to spend about $220 billion, and on 23 October Andy Jassy said AWS capacity for 2026 was sold out and 2027 largely sold out.
Since GTC, NVIDIA has opened its Vera Rubin rack designs to the Open Compute Project. Google Cloud has set a launch date for its Vera Rubin A5X instances. The test of the $1tn figure is whether the buyers’ budgets keep rising through 2027.
Sources
- CNBC: Nvidia GTC 2026: CEO Jensen Huang sees $1 trillion in orders for Blackwell and Vera Rubin through ’27, 16 March 2026
- NVIDIA Blog: GTC 2026 live updates
- NVIDIA Newsroom: NVIDIA Vera Rubin Opens Agentic AI Frontier
- SEC EDGAR: Meta Reports Fourth Quarter and Full Year 2025 Results, Exhibit 99.1, 28 January 2026
- Meta Investor Relations: Meta Reports First Quarter 2026 Results
- Alphabet Investor Relations: 2025 Q4 Earnings Call
- CNBC: Amazon stock falls 8% on $200 billion spending forecast, earnings miss, 5 February 2026
- CNBC: Alphabet resets the bar for AI infrastructure spending, 4 February 2026
- CNBC: Meta stock climbs on report of planned layoffs, 16 March 2026
- SEC EDGAR: Nebius Group Form 6-K, commercial agreement with Meta, 16 March 2026
- Nebius: Nebius signs new AI infrastructure agreement with Meta, 16 March 2026
Related stories
- NVIDIA GTC 2026: The $1 Trillion Infrastructure Bet Reshaping AI
- Metas 27 Billion Nebius Gamble: Outsourcing the Future of AI
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
