OpenAI now runs pieces of ChatGPT on Google’s Tensor Processing Units, the custom chips built by Microsoft’s most important AI rival. Sam Altman’s company still calls Microsoft Azure its primary home. But its own infrastructure choices are quietly rewriting the Alphabet versus Microsoft debate that Wall Street keeps reducing to a spreadsheet.
Investors comparing Gemini’s ad growth against Copilot’s seat count are missing a cleaner signal. OpenAI and Anthropic, the two labs that matter most in generative AI, are voting with their own infrastructure budgets, and Google keeps cashing the check.
OpenAI Quietly Becomes a Google Cloud Customer
From 2019 through early 2025, Microsoft Azure had exclusive rights to host every OpenAI workload. That exclusivity ended in January 2025, opening the door for OpenAI to rent capacity elsewhere. It didn’t take long to walk through it.
OpenAI has started running ChatGPT and other products on Google’s TPUs, marking the first time it has leaned on chips beyond Nvidia’s GPUs at real scale. Google has kept a boundary in place. It has not handed OpenAI access to its most powerful TPU models, a restriction one Google Cloud employee confirmed to reporters. Even limited, the arrangement is a signal aimed straight at Redmond: OpenAI is willing to lean on a direct Microsoft competitor for infrastructure.
Microsoft and OpenAI have since rewritten their own commercial terms too. The amended agreement, announced this spring, changed several load-bearing pieces of the original partnership at once.
- Microsoft no longer pays OpenAI a revenue share when customers reach OpenAI’s models through Azure
- OpenAI keeps paying Microsoft a 20 percent cut through 2030, but that obligation now carries a total cap
- Microsoft’s license to OpenAI’s models and products runs through 2032, but it is no longer exclusive
- OpenAI can now sell its products through any cloud provider, including Amazon and Google
Each change trades away a piece of Microsoft’s original leverage in exchange for flexibility on both sides. Microsoft still gets to call OpenAI a partner. It just cannot call it a captive one anymore.
A Decade-Old Chip Bet Starts Collecting Rent
Alphabet’s own numbers already look strong without any of this. Revenue rose 22% in the first quarter of 2026, Google Cloud revenue jumped 63%, and cloud operating profit more than tripled to $6.6 billion. Gemini’s integration across Search, Android, and Google’s software lineup gets most of the credit for that in the usual telling.
The chips underneath that cloud business have their own decade-long backstory. Google’s TPU program began in 2013, first shipped in 2015, and is now on its seventh generation, code-named Ironwood, which became generally available in 2025. That head start over Microsoft and Amazon’s own custom-silicon efforts shows up in a widening list of outside customers who now pay to use it.
| Customer | Deal Terms | Primary Cloud Status |
|---|---|---|
| Anthropic | Deal worth up to $40 billion covering five gigawatts of TPU capacity over five years, with access to as many as one million Ironwood chips | Still names Amazon Web Services its primary training partner |
| Meta | Six-year cloud agreement worth more than $10 billion, signed in 2025 | Runs its own custom chips alongside Nvidia GPUs |
| OpenAI | Renting TPU capacity through Google Cloud since mid-2025, without access to Google’s top-tier TPU models | Azure remains primary; new products still ship there first |
| Apple | Used TPUs for on-device AI model training starting in 2024 | No reported cloud exclusivity commitments |
Anthropic’s own announcement of a $40 billion, five-gigawatt TPU commitment is the clearest external validation Google has landed to date. The company still trains and runs its Claude models across three different chip families for redundancy, but the fact that it is willing to hand Google that much money is its own kind of review. On raw economics, Google’s chips reportedly carry a total cost 44% below Nvidia’s GB200 servers for a comparable configuration, according to semiconductor analysis firm SemiAnalysis. That is the kind of gap that gets a rival’s own partner to at least test the waters.
Google’s Own Scientists Are Waiting for Chips
Success like that has a cost Google did not fully plan for. Bloomberg reported that Google’s own AI researchers, including teams inside Google DeepMind, are now competing internally for the same computing capacity the company is selling to Anthropic and Meta.
DeepMind chief executive Demis Hassabis has acknowledged the squeeze publicly, pointing to what he called “a few suppliers of a few key components.” It is a polite way of describing the high-bandwidth memory bottleneck at Samsung, Micron, and SK Hynix that caps how fast any company can build advanced accelerators, TPUs included. That same memory crunch helped trigger a chip rout that erased $1 trillion in tech value across markets in Seoul and New York earlier this year.
Researchers need volume to experiment, not just access. When the chips are already booked by paying customers, experiments wait. At least one long-tenured DeepMind researcher, Ioannis Antonoglou, has left for a startup role in the past 18 months, part of a pattern that has picked up as internal compute access has tightened.
That is the irony sitting inside Alphabet’s chip story. The same TPU stack that is winning outside customers and validating a decade of spending is now rationed even for the people who built it.
How Much Is Microsoft’s Consolation Prize Worth?
Microsoft is not losing this fight by any conventional measure. Its roughly 27% stake in OpenAI is worth around $230 billion, its commercial remaining performance obligations across the whole company doubled to $627 billion, and Azure keeps growing off the back of ChatGPT’s usage. The scoreboard just looks different once you get past the top line.
- $230 billion – the estimated value of Microsoft’s roughly 27% stake in OpenAI
- 420 million – Copilot’s monthly active users in the first quarter of 2026, up from 230 million a year earlier
- 11.5% – Copilot’s share of paid AI subscribers in January 2026, behind ChatGPT’s 55.2% and Gemini’s 15.7%, per Recon Analytics tracking
- 35.8% – the share of employees with Copilot access who actually use it on a regular basis, according to industry adoption surveys
Scale and engagement are pulling in opposite directions. Copilot has genuine distribution, riding inside Microsoft 365 seats companies already own, yet it still trails both ChatGPT and Gemini in the market for paid AI subscriptions it was built to win.
The capped revenue share may hurt OpenAI down the road, and gives Microsoft a way to not overpay.
That is Holger Mueller, an analyst at Constellation Research, describing a capped revenue share that could hurt OpenAI under the amended terms. His read cuts both ways: Microsoft protected itself from a runaway payout, but it also gave up the exclusivity that made the original bet so lopsided in its favor.
The Capex Bill Keeps Climbing for Both Companies
None of this infrastructure race is cheap, and the bill is starting to show up in the filings. Alphabet spent $35.7 billion on capital expenditures in the first quarter of 2026 alone, according to capital spending that more than doubled to $35.7 billion in its own quarterly filing, up from $17.2 billion a year earlier. Full-year guidance now sits at $175 billion to $185 billion, nearly double what the company spent in 2025.
Microsoft is running at its own annualized pace of roughly $145 billion for its 2026 fiscal year. Combined with Amazon and Meta, the four largest hyperscalers are on pace for combined 2026 capital spending near $725 billion, up roughly 77% from about $410 billion in 2025.
Spending that fast raises an old question in a new setting: what happens to earnings once all that hardware starts depreciating. IBM’s 25% stock crash tied to a 2014 bet is a reminder that infrastructure decisions made in one cycle can boomerang years later, long after the headlines about them have faded. Alphabet and Microsoft are both betting their AI infrastructure ages better than that.
Valuation Still Splits the Difference
Strip away the chip subplot and the two stocks still land close together on paper. Wall Street analysts expect both companies to grow earnings by roughly 16% to 17% annually over the next three to five years. Microsoft trades at about 20 times forward earnings, versus roughly 24 times for Alphabet.
That gap is the entire case for Microsoft as the cheaper way to own the same AI buildout, a case laid out in more detail in the wider Microsoft-Alphabet AI stock matchup. It is a real argument. Cheaper earnings on a similar growth path is not nothing.
But the infrastructure story adds a layer that multiple alone will not capture. Alphabet’s TPU stack is now good enough that Microsoft’s own former exclusive partner rents it, and good enough that Google’s own researchers have to wait their turn behind paying customers. Microsoft’s OpenAI stake is still worth a fortune. It is just a smaller and less exclusive fortune than it looked eighteen months ago.
Frequently Asked Questions
What is a TPU and why does it matter to Alphabet’s stock?
A TPU, or Tensor Processing Unit, is Google’s own chip built specifically for AI math rather than general computing. Google’s seventh-generation Ironwood chip claims roughly 4.7 times better price-performance and 67% lower power draw than comparable high-end Nvidia GPUs on many AI workloads, which is why outside labs increasingly pay to rent it.
Did Microsoft’s stake in OpenAI change under the amended deal?
No. The equity position stayed intact. Only the commercial terms changed, including cloud exclusivity and revenue sharing; Microsoft still holds intellectual property rights to OpenAI’s models through 2032, just on a non-exclusive basis now.
Is Nvidia losing business to Google’s chips?
Not yet in a way that shows up in its results. Nvidia’s fiscal 2026 revenue, for the year ended in January, came in near $216 billion, with data center sales alone around $197 billion. Google’s TPU customers remain a small slice of the overall AI chip market even as their number grows.
Why did OpenAI diversify away from Microsoft Azure in the first place?
Cost and redundancy. Beyond Google Cloud, OpenAI has expanded computing deals with Oracle and CoreWeave and signed a SoftBank-backed data center project reported to be worth around $500 billion, spreading its infrastructure risk beyond any single partner, Microsoft included.
Could Google Cloud overtake Azure in size?
Not soon, based on current figures. Google Cloud’s own contract backlog sits near $240 billion, while Microsoft’s total commercial backlog across every product line doubled to $627 billion. Google is growing faster in percentage terms, but from a considerably smaller base.
Disclaimer: This article is for informational purposes only and does not constitute investment advice. Stock ownership carries the risk of loss, figures are accurate as of publication in July 2026, and readers should consult a licensed financial advisor before making investment decisions.
