Amazon Bets $50 Billion on OpenAI While NVIDIA Drops the Vera Rubin Roadmap — The Real AI War Is About Infrastructure

Amazon CEO Andy Jassy's framing was straightforward. He wrote on LinkedIn that Amazon believes OpenAI "will be one of the big winners in AI" and that Amazon "expects to earn a strong long-term return." Analysts at William Blair put numbers on that optimism: the deal could generate $17 billion in annual revenue for AWS, roughly 11% of its expected 2026 revenue. [1] Sam Altman, for his part, said the partnership "helps put powerful AI into the hands of businesses at real scale." Translation: OpenAI needs infrastructure it can't build alone, and Amazon has more of it than almost anyone on Earth.

The Stargate collapse and Microsoft's legal threat

Amazon's timing wasn't accidental. The Stargate project — a planned $500 billion data center consortium between OpenAI, Oracle, and SoftBank — has reportedly fallen apart after negotiations with Oracle broke down. Stargate was supposed to be the infrastructure backbone for OpenAI's most ambitious AI plans. Its collapse left a vacuum, and Amazon moved to fill it. [1] This puts Microsoft in an awkward position. Microsoft invested $1 billion in OpenAI in 2019 and another $10 billion in 2023. Under their agreement, all access to OpenAI's models was supposed to be routed through Microsoft's Azure cloud platform. That exclusivity has driven record revenues for Azure. [5]

Now Amazon is building something called a "stateful runtime environment" on Amazon Bedrock — a system that lets developers keep context, remember prior work, and operate AI agents across software tools. OpenAI and Amazon argue this is fundamentally different from the "stateless" API access that Microsoft controls. Microsoft's lawyers apparently disagree. [5] The Financial Times reports that Microsoft is considering legal action, arguing the Amazon deal violates the spirit — if not the letter — of its exclusive cloud agreement with OpenAI. Whether "Frontier" (OpenAI's new enterprise platform for AI agents, distributed exclusively through AWS) counts as stateful or stateless access could reshape cloud alliances across the entire AI industry. [5]

Microsoft said publicly that it is "confident that OpenAI understands and respects the importance of living up to its legal obligations." OpenAI and Amazon declined to comment. For now, all three companies are talking, hoping to avoid a courtroom fight that nobody wants but somebody might need. [5] The strategic read is clear: OpenAI is diversifying away from its dependence on any single cloud partner. Microsoft made that possible by funding OpenAI when nobody else would. Amazon is making it lucrative by offering infrastructure at a scale that even Azure can't easily match. And OpenAI is playing them against each other — the same way every growing company eventually outgrows its first investor.

NVIDIA's answer to the skeptics

While Amazon was rewriting the cloud computing map, Jensen Huang walked onto the stage at GTC 2026 in San Jose and spent two hours making one argument: the AI buildout is not slowing down. NVIDIA had a credibility problem heading into the keynote. The company posted $68.1 billion in quarterly revenue — a record — and watched its stock fall 14% in the aftermath. Investors had already priced in strong growth. What they wanted to know was whether AI infrastructure spending would keep compounding or start to level off. [3]

Huang's answer was the Vera Rubin platform, named after the astronomer whose work revealed dark matter. Each Vera Rubin GPU delivers 50 petaflops of NVFP4 inference performance — five times what Blackwell could do. At the rack level, the NVL72 configuration puts out 3.6 exaflops of inference compute and 700 million tokens per second. The previous generation managed 22 million tokens from an entire gigawatt data center. [3][4] The numbers are almost absurd. Huang walked through them with the careful pacing of someone who knows his audience has calculators. Moore's Law would have delivered maybe 1.5x improvement in the same timeframe. NVIDIA delivered 350x. [4]

But the Vera Rubin GPU wasn't the only thing on stage. Huang also unveiled the Rubin Ultra NVL576 — a system arriving in the second half of 2027 that packs 576 GPU chiplets, 365 terabytes of HBM4e memory, and 15 exaflops of inference compute. The power envelope: 600 kilowatts per rack. That's not a typo. A single rack consuming enough electricity to power a small neighborhood. [3] And then, looking even further ahead, Huang showed the first preview of Feynman — NVIDIA's 2028 architecture built on TSMC's 1.6-nanometer A16 process node, the most advanced semiconductor manufacturing technology ever brought to mass production. Feynman is designed as an "inference-first" chip, built specifically for the long-context, multi-step reasoning that AI agents need. It also introduces silicon photonics — replacing copper electrical connections with optical signals for chip-to-chip communication. [3]

That last detail matters more than it might seem. At the scales NVIDIA is operating, electrical interconnects are hitting physical limits. Moving to light-based communication isn't incremental improvement. It's a wholesale change to how AI infrastructure works at the most fundamental level.

The trillion-dollar claim

Huang made one number the centerpiece of his keynote: $1 trillion. That's what he says NVIDIA can see in combined Blackwell and Vera Rubin orders through 2027. Last year at GTC, that figure was $500 billion. It doubled in twelve months. [4] "Does it make any sense?" Huang asked the crowd of 30,000. Then he spent the rest of the keynote trying to prove that it does.

The logic goes like this: over 90% of NVIDIA's revenue now comes from AI data centers. Not gaming, not professional graphics — AI infrastructure. Every major cloud provider (AWS, Azure, Google Cloud, Oracle, CoreWeave) is placing massive orders. Meta, Alibaba, and dozens of "AI native" startups are adding demand on top of that. NVIDIA says it runs every domain of AI across every type of AI model, and between NVIDIA, Anthropic, and Meta, that represents roughly a third of the world's AI compute. [3][4] The skeptics have reasonable concerns. NVIDIA has a history of announcing extraordinary benchmark numbers at conferences that prove difficult to replicate in production. The 50 petaflops figure applies to NVFP4 precision — a data format that not every model or inference pipeline can use. The "10x cost per token" improvement applies specifically to mixture-of-experts model inference, not dense models that many organizations still run. [3]

Bank of America maintained a $300 price target heading into GTC. The market consensus sat around $273. Whether the AI infrastructure buildout can sustain NVIDIA's growth trajectory into 2027 and beyond is a question Huang answers confidently every year. Whether the market believes him is a different story.

The real bottleneck isn't chips

Here's what both the Amazon deal and the NVIDIA roadmap are telling you, if you read between the lines: the AI race has shifted. Two years ago, the constraint was compute. Companies couldn't get enough GPUs. NVIDIA had a supply problem, and everyone was waiting in line. That bottleneck is easing. Vera Rubin is already in production. AWS has committed to over a million NVIDIA GPUs globally. The chips are coming. [3][4]

The new bottleneck is everything else. Power. Cooling. Physical space. Permitting. Water rights. The ability to build data centers fast enough to house all the silicon that's being manufactured. Amazon's $200 billion in 2026 capital expenditure isn't mostly going to chips. It's going to land, concrete, power infrastructure, cooling systems, and the engineering required to make a building that consumes city-level electricity actually function. The company's €33.7 billion investment in Spain's data center infrastructure tells the same story — this is a construction problem now, not a semiconductor problem. [1]

NVIDIA's Rubin Ultra rack consuming 600 kilowatts makes this concrete. Data centers built for today's AI workloads were not designed for 600kW per rack. Retrofitting existing facilities — or building new ones — adds years and billions in capital costs that don't show up in chip benchmark comparisons. [3] TSMC's position amplifies the complexity. The Taiwanese chipmaker now manufactures chips for NVIDIA, Apple, AMD, and most of the AI industry. NVIDIA's Feynman architecture is built on TSMC's most advanced 1.6-nanometer process. That makes Taiwan — a 14,000-square-mile island 100 miles from mainland China — arguably the most strategically important piece of real estate on Earth. Every AI data center in the world depends on chips that largely come from one place. [3]

What this means for everyone else

The Amazon-OpenAI deal and NVIDIA's GTC roadmap are two chapters of the same story: AI is becoming an infrastructure war, and the winners will be determined by who can build the biggest, most efficient factories to produce intelligence at scale. For companies trying to deploy AI: inference costs are about to drop significantly. NVIDIA is promising 10x lower cost per token with Vera Rubin. That changes the economics on automation use cases that were borderline viable in 2025. Workflows that didn't justify the compute spend will make obvious financial sense by the end of 2026. [3]

For cloud providers: the power dynamic is shifting. OpenAI playing Amazon and Microsoft against each other is the opening move in what will become a broader pattern. As AI models become commoditized, the cloud infrastructure they run on becomes the real competitive moat. AWS, Azure, and Google Cloud are all spending hundreds of billions to make sure they're the ones collecting the toll. For investors: the question isn't whether AI spending is real. Amazon wouldn't commit $50 billion and $200 billion in capex to a bubble. NVIDIA wouldn't be showing 2028 chip roadmaps if it expected demand to evaporate. The question is whether the returns on this spending show up in corporate earnings fast enough to justify the valuations — or whether we're watching the most expensive infrastructure buildout in history outrun its own business case.

Jensen Huang looked at 30,000 people in San Jose and told them the AI buildout isn't even close to done. Amazon's checkbook says the same thing. Whether they're right will determine the trajectory of the technology industry for the next decade.