Meta's TPU Deal Sent Nvidia Sliding on Wall Street

On June 18, Nvidia shares fell 2.6% after The Information reported that Google is in talks with Meta Platforms to spend billions of dollars installing Google's tensor processing units in Meta's own data centers as soon as 2027. Advanced Micro Devices dropped more than 4% on the same news. The move was not a surprise to anyone who had read the February headlines — Meta had already signed a multiyear, multibillion-dollar deal to rent Google's Ironwood TPUs through Google Cloud for training next-generation Llama models. What changed in June was the escalation: Google is no longer content to rent chips from its own facilities. It is pitching hyperscalers to buy TPUs outright and install them beside Nvidia GPUs and AMD Instinct accelerators. Meta, which told investors in January to expect $115 billion to $135 billion in AI infrastructure spending in 2026, is now hedging across every major silicon vendor on the planet. Wall Street's reaction was modest in percentage terms. In strategic terms, it was enormous.

The February deal sequence tells you how deliberate this is. On February 17, Meta expanded its Nvidia partnership to buy millions of Blackwell and Rubin GPUs plus Grace CPUs and Spectrum-X networking — tens of billions of dollars committed. One week later, on February 24, Meta struck a deal with AMD worth up to $100 billion for six gigawatts of Instinct MI450 GPUs, including a warrant for 160 million AMD shares. On February 26, The Information broke the Google TPU rental agreement. Meta did not defect from Nvidia. It bought everything available. That is not a company shopping for a cheaper vendor. It is a company that has concluded no single supplier can meet frontier demand — and that telling Nvidia it has alternatives is itself a negotiating weapon. The Decoder reported that OpenAI secured roughly 30% lower prices from Nvidia simply because Google's TPUs existed as a credible outside option. Pricing power, not market share, is what breaks first.

Google's commercial push is what makes this cycle different from every prior TPU headline. For a decade, Google's tensor processing units were an internal advantage — powering YouTube recommendations, then Gemini, but rarely sold as a merchant product. That changed in May 2026, when Google announced it would sell TPUs directly to customers for the first time and unveiled an inference-specialized chip aimed at Nvidia's Groq LPU. Amin Vahdat, promoted in December to chief technologist for AI infrastructure reporting directly to CEO Sundar Pichai, is running the offensive. Google Cloud executives have privately set a target of capturing as much as 10% of Nvidia's annual revenue through expanded TPU sales — roughly $20 billion at current run rates, according to reporting in The Decoder and SiliconANGLE. Google is also forming a joint venture with an institutional partner to lease TPUs to outside customers, and struck a $5 billion deal with Blackstone to launch a cloud-services company competing with Nvidia-backed CoreWeave and Nebius.

The financing layer is where Google stops competing on chips and starts competing on industrial strategy. Lumida News reported on June 19 that Google is providing multibillion-dollar financial guarantees to data-center developers who commit to TPU deployments — including a $3.2 billion guarantee for the Lake Mariner cluster in western New York, where Anthropic will rent compute; a $7 billion project called River Bend near Baton Rouge; and $1.4 billion for a facility in Colorado City, Texas. This is Nvidia's own playbook: stoke demand by backstopping the buildout, then capture the silicon that fills the racks. Google has raised $85 billion in equity to fund its AI infrastructure expansion — more financial firepower than any prior Nvidia challenger has brought to the fight. The fear that cloud providers call "Jensen jail" — losing Nvidia GPU allocations if you buy from a rival — has been the primary brake on TPU adoption. Blackstone's willingness to partner with Google despite its ties to Nvidia-backed CoreWeave suggests that brake is slipping.

Nvidia's public response has been confident to the point of defensiveness. After the June report, a company spokesperson told Yahoo Finance: "We're delighted by Google's success — they've made great advances in AI and we continue to supply to Google." The company added that "NVIDIA is a generation ahead of the industry" and "the only platform that runs every AI model and does it everywhere computing is done." CEO Jensen Huang, on an earlier earnings call, noted that Google remains a Nvidia customer and that Gemini can run on Nvidia hardware. He disclosed visibility into $500 billion in revenue from Blackwell and Rubin platforms through 2026. Over the same weekend, Nvidia sent Wall Street analysts a memo — obtained by Yahoo Finance — rejecting comparisons to Enron, WorldCom, and Lucent, and insisting its strategic investments are transparent and its underlying business economically sound. The memo was a response to Michael Burry, who has publicly bet against the company and argued the AI market resembles the dot-com bubble. When a monopolist has to issue an accounting-defense memo and a competitor-success congratulation in the same week, the monopoly narrative is already cracking.

The market data supports a more nuanced story than "Nvidia is finished." Analysts estimate Nvidia still holds 75% to 90% of AI accelerator revenue, defended by CUDA's software moat and plug-and-play hardware integration. But custom silicon from Google, Amazon, Microsoft, and Meta collectively accounts for an estimated 15% to 20% of the accelerator-server market and is growing roughly 45% year over year — faster than merchant GPU shipments. The total pie is expanding: hyperscaler capex is projected above $650 billion in 2026, up sharply from 2025. Meta's three-vendor strategy is a bet that the pie grows faster than any single vendor's share — not that Nvidia disappears. DA Davidson estimated in September that Google's TPU business and DeepMind together could be worth $900 billion. Citadel Securities has reported running key workloads 30% cheaper and four times faster on TPUs, according to Lumida News. The performance gap that once made TPUs a curiosity is closing at the exact moment inference — not training — becomes the dominant cost center.

The timing intersects brutally with the AI IPO calendar. OpenAI filed confidentially on June 8. Anthropic filed June 1. SpaceX completed its first full trading week as a public company the same week Nvidia sold off on the Meta-Google report. Oracle — which beat earnings, posted a $638 billion AI backlog, and pledged $70 billion in capex — lost roughly a quarter of its market value anyway, as Convina reported earlier today. The market is no longer rewarding infrastructure ambition without questioning who finances it and who captures the returns. Nvidia sits at the center of that question: it is simultaneously the indispensable supplier, a strategic investor in its own customers, and the company whose largest buyers are building credible alternatives. Mizuho analysts, reacting to the TPU news, noted that a shift toward Google chips is positive for Broadcom and Lumentum's optical compute switches — and "a modest challenge" for GPU suppliers. Modest, for now. The precedent is not.

Convina's view: Nvidia's moat was never just chips. It was the credible threat that if you did not buy Nvidia, you did not get Nvidia — and in a supply-constrained market, that was everything. Google's decision to sell TPUs directly, finance data centers the way Nvidia financed its ecosystem, and land Meta as a flagship external customer attacks that logic at the procurement layer, not the benchmark layer. Meta is not leaving Nvidia. It is making Nvidia compete. That is a harder problem for a company trading as a monopoly than for one trading as a leader. Enterprise buyers should read this week as a procurement signal: multi-vendor AI infrastructure is no longer a contingency plan. It is the default architecture for any organization spending at frontier scale. The companies that locked in single-vendor contracts in 2024 are holding agreements written for a market that ended in 2026. Nvidia may remain a generation ahead on silicon. But the generation that mattered was the one where one vendor controlled the stack. That generation is over — and Wall Street, finally, is pricing it.