When Google also wants to “print shares” to build AI, whose move broke Neocloud’s high-valuation narrative?

Author: Ada, Deep Tide TechFlow

Recently, Google announced its first equity financing since 2005. Looking at three actions by Google over the past 90 days in sequence, this $80 billion figure is not just about capacity issues—it points to the very premise that Nvidia GPUs are set to dominate the entire AI compute market. The most directly affected are the Neocloud “trio” who have priced their valuations by betting on “Nvidia exclusivity”: CoreWeave, Nebius, and IREN.

The complete puzzle formed by three moves

On April 22, at Google Cloud Next ’26, Google released the eighth-generation TPU. It was split into two chips: TPU 8t dedicated to training and TPU 8i dedicated to inference. In the same product announcement, Google for the first time stated explicitly that it will sell TPU externally to selected third-party data center operators. This is the first time in ten years since TPU entered mass production in 2015 that it has officially been offered beyond Google Cloud.

On May 24, Google and Blackstone announced the formation of a joint venture. Blackstone initially injected $5 billion in equity; after leverage, the total scale could reach $25 billion. With Blackstone as the majority shareholder, Google would provide TPU and software. The new company is positioned as a compute-as-a-service provider—precisely the standard operating model of Neocloud. The target is to deploy 500 MW of capacity by 2027, led by Benjamin Treynor Sloss, a former Google executive. On the day of the announcement, CoreWeave fell 3.8% and Nebius fell 1%.

On June 1, Google announced an $80 billion equity financing. It filled to the maximum a once-in-a-decade equity instrument that had never been touched since 2005: $15 billion in convertible preferred stock, $15 billion underwriting issuance of A/C class common stock, a $40 billion on-exchange ATM follow-on issuance plan, and a $10 billion Buffett private placement.

Taken together, these three actions show that Google is laying three paths at the same time: building data centers in-house, selling chips, and building Neocloud. In essence, these are three penetration modes of the same TPU compute stack outward. Calling it a big-tech expansion vastly underestimates Google’s ambition. It is trying to remake the Nvidia GPU-dominated compute market with TPU.

The real reason behind the $80 billion equity financing

Media releases attribute this financing entirely to AI infrastructure, which is a misread. Google itself lays it out clearly in its SEC filings: within the $40 billion ATM plan, roughly $30 billion is intended to cover tax obligations related to employee equity incentives in 2026. That qualifies as an “administrative arrangement,” not new capital expenditure.

After removing this portion, the true “new money” used for AI infrastructure is about $50 billion: $15 billion underwriting issuance plus the $10 billion Buffett private placement plus the remaining $15 billion from the ATM.

Comparing to another number, Google’s full-year 2026 capital expenditure guidance is $180–190 billion, and it “will significantly increase” in 2027. An $80 billion equity financing can only cover a bit more than one-quarter of annual capital expenditures; the rest must be filled through operating cash flow, debt, and subsequent financing.

This, in turn, explains why Google has to move its equity. Google Cloud’s Q1 2026 revenue rose 63% year over year, and backlog orders doubled from $230 billion in the previous quarter to more than $460 billion. Even demand coming in through signed contracts alone already far exceeds the pace at which Google can expand its self-built capacity. In other words, even a cash cow like Google has reached the point where AI capital expenditures are so large that it must start diluting equity.

Berkshire Hathaway’s $10 billion private placement is another detail in this financing that needs to be separated and examined on its own. Over the past 60 years, Buffett’s public record shows that he almost never participates in primary markets, and certainly not in equity/capital-expenditure financing for “new economy” companies. This time, with A-class priced at $351.81 per share and C-class at $348.20 per share—fixed-price purchases—it looks more like a form of identity certification: stamping “AI compute as an infrastructure asset class.”

Microsoft’s route vs. Google’s route—now diverging

To understand the true meaning of this financing, you need to put the two largest compute buyers side by side.

Microsoft follows the “self-build plus Neocloud outsourcing” route. Its in-house chip, Maia, has fallen short of expectations, and compute demand for OpenAI’s training and inference has grown exponentially. Since the end of 2025, Microsoft’s contract commitments to Neocloud have exceeded $60 billion: Nscale at $23 billion (for deploying 200K GB300 chips), with the rest allocated among CoreWeave, Nebius, IREN, and Lambda Labs. These contracts are all for Nvidia GPU-based deployments. Microsoft has to rely deeply on Neocloud because its own built capacity can’t keep up with demand, and its self-developed chips can’t match Nvidia.

Google is taking a different path. It develops TPU in-house, builds data centers independently (not relying on Neocloud), and now it sells TPU to others as well, while using the Blackstone JV to compete in the Neocloud market. Google doesn’t need Neocloud; it intends to become a competitor to Neocloud itself.

This divergence is the strategic pivot behind this financing. The more deeply Microsoft binds itself to Neocloud, the more Google needs to break Neocloud. The two companies’ choices differ because their underlying assets are different: Microsoft does not have its own high-end AI chips, while Google has TPU.

What supports Google’s path is the real progress of TPU. By 2025, Anthropic had already migrated large-scale training tasks onto TPU. Meta, SSI, and xAI have also been reported to be negotiating TPU orders. Google’s internal view is that TPU’s cost-performance ratio for specific inference workflows is 3 to 5 times that of Nvidia GPUs, and this figure has been validated by multiple independent analysts.

The asymmetric fate of the trio

Looking back at the Neocloud trio: CoreWeave, Nebius, and IREN.

In the short term, Google does not pose a threat to them. CoreWeave’s backlog orders reached nearly $100 billion in Q1, including Meta’s $21 billion newly signed contract in March and multi-year contracts with Anthropic. Nebius’s Q1 revenue was $390 million, up 841% year over year. Its full-year guidance is $3–3.4 billion, with an annualized operating rate of $7–9 billion, and it has a five-year contract with Meta worth $27 billion. IREN holds $9.7 billion in contracts with Microsoft and $5.5 billion in contracts with Nvidia. These are all locked-in Nvidia GPU contracts that Google TPU cannot replace.

What is being broken is the valuation narrative. The logic behind the three companies’ high valuations rests on three premises: AI compute demand is extremely tight, Nvidia GPUs are the only option, and hyperscalers’ self-build efforts can’t keep up with demand. Google’s combined moves are loosening each of these premises one by one. TPU is a real alternative, new capacity is catching up, and self-build can’t keep pace—so JV-backed acceleration becomes the solution.

But the three companies’ situations are completely different.

CoreWeave: its high-valuation risk has been partially released, but its debt leverage has not been cleared. Its market positioning is “AWS in the GPU era,” with the highest ambitions and the highest valuation premium. Nvidia already holds about 11% equity in CoreWeave, with a market value close to $4.9 billion, and it doubled its stake in January 2026 at $87.20 per share. This deep linkage leaves CoreWeave no room to pivot to TPU, because in customers’ perceptions it is the agent of Nvidia GPUs. Google’s strategy only needs to persuade the market that TPU is truly a front-line option, and CoreWeave’s valuation premium will shrink.

Nebius: it sits in the middle. Its technology stack is relatively open (Soperator has been open-sourced, similar to CoreWeave’s SUNK route). Although its customer structure is still Nvidia GPU-heavy, it has more flexibility. Nebius’s debt and cash are close to offsetting each other. In late May, after Google entered, Situational Awareness—an investment fund run by former OpenAI researcher Leopold Aschenbrenner—started building a position in Nebius. His bet is on which will run faster: revenue growth or valuation growth.

IREN: the most unusual case. This company transitioned from Bitcoin mining into AI infrastructure; among the trio, it has the heaviest asset base and the lowest valuation premium. The cash flows from contracts with Microsoft ($9.7 billion) and Nvidia ($5.5 billion) are enough to sustain its fundamentals. It faces no pressure from a “high-valuation narrative” being broken. In the new landscape, IREN has shifted from “the weakest” to “the most stable,” but it is also not cheap anymore.

From a compute market of shortage to customer stratification

The second-order implication of all this is a structural change in the compute market.

In the past 18 months, the AI compute market was a typical sellers’ market. Nvidia determined the supply cadence, and every buyer queued up. Now, three layers are happening simultaneously.

First, frontier model labs are starting to become multi-stack. Anthropic has already publicly used Google TPU, AWS Trainium, and Nvidia GPUs, and OpenAI has also been reported to be evaluating TPU. Once multi-stack becomes the standard configuration for leading labs, the “exclusive Nvidia GPU” Neocloud label becomes, from the customer’s perspective, a limitation.

Second, hyperscaler routes are splitting. Microsoft (deeply tied to Neocloud), Google (self-build plus selling chips plus building Neocloud), and Amazon (mainly in-house Trainium development) are taking completely different directions. This split directly determines Neocloud’s customer structure. Currently, Neocloud’s key customers are Microsoft and Meta; Google is completely absent. If Microsoft were to reduce outsourcing due to Maia improvements or adjustments in its relationship with OpenAI, Neocloud’s revenue side would face structural risk.

Third, the cost of capital is stratifying. Google funds itself with equity plus Buffett-backed credibility plus operating cash flow, keeping the cost of capital close to zero. CoreWeave’s latest loan pricing is SOFR (secured overnight financing rate) + 4.5%. In a capital-intensive business where GPUs have a depreciation cycle of only 5 to 7 years, this cost-of-capital gap compounds into a deadly difference. Neocloud can exist only because Nvidia GPUs are still hot commodities; once GPUs shift from scarce items to relatively abundant goods, the players with the lowest funding cost will regain dominance in the market. That is the direction Google is betting on.

Track the next three indicators

Returning to that $80 billion equity financing, the real signal it sends to the market is that Google is treating AI compute as a market to be re-divided and is preparing to fight for it. In the short term, the contracts of CoreWeave, Nebius, and IREN can still run for 2 to 3 years, but the “Nvidia exclusivity” belief underpinning their high valuations has already been cracked open by Google’s combined moves from the outside.

From here, tracking three things is enough: whether the Google–Blackstone JV can light up the 500 MW capacity on schedule in 2027; whether the TPU customer list can expand from Anthropic to Meta and xAI; and whether Microsoft will circle back to negotiate TPU if tensions arise in the OpenAI relationship. If any two out of these three come through, the story of the trio will need to be rewritten.