Author: Ada, Deep Tide TechFlow

Recently, Google announced its first equity financing since 2005. Looking at the three actions Google has taken in the past 90 days, this 80 billion is not just about capacity issues; it points to the fact that Nvidia GPU dominates the entire AI computing power market. The most directly impacted are the Neocloud trio betting on "Nvidia uniqueness": CoreWeave, Nebius, IREN.

The complete picture with three actions combined

On April 22, at the Google Cloud Next '26 conference, Google released the eighth-generation TPU, differentiating into TPU 8t for training and TPU 8i for inference. In the same product announcement, Google explicitly stated for the first time that it will sell TPUs to selected third-party data center operators. This is the first official departure of TPUs from Google Cloud in ten years since mass production began in 2015.

On May 24, Google announced a joint venture with Blackstone. Blackstone's initial investment is 5 billion USD in equity, which can scale up to 25 billion with leverage, with Blackstone serving as the major shareholder, and Google providing TPUs and software. The new company is positioned as a compute-as-a-service provider, which is exactly the standard business model of Neocloud. The goal is to deploy 500 megawatts of capacity by 2027, led by former Google executive Benjamin Treynor Sloss. On the day of the announcement, CoreWeave dropped by 3.8%, and Nebius fell by 1%.

On June 1, Google announced an 80 billion USD equity financing. It filled its unused equity tool for the first time since 2005, including 15 billion in convertible preferred stock, 15 billion in underwriting issuance of A/C class common stock, a 40 billion on-market ATM issuance plan, and 10 billion from Buffett's private placement.

Looking at the three actions together, Google is simultaneously laying out three paths: building its own data centers, selling chips, and creating Neocloud. These are essentially three forms of penetration for the same TPU computing power stack. To say this is merely a big tech company expanding production seriously underestimates Google’s ambitions. It is attempting to remake the Nvidia GPU-dominated computing power market with TPU.

The real reason for the 80 billion equity financing

Media reports erroneously attributed this financing entirely to AI infrastructure. Google clearly stated in its SEC filings that approximately 30 billion of the 40 billion ATM plan is intended to cover 2026 employee stock incentives’ tax obligations, which is an "administrative arrangement" rather than new capital expenditure.

Excluding this part, the real "new money" for AI infrastructure is about 50 billion, which includes 30 billion for underwriting issuance, 10 billion from Buffett's private placement, and 10 billion remaining from the ATM.

In comparison to another figure, Google’s capital expenditure guidance for the entire year of 2026 is between 180 billion and 190 billion USD, and it is expected to "increase significantly" in 2027. The 50 billion equity financing can only cover a little over 1/4 of the annual capital expenditure, with the remaining funds needing to be filled by operating cash flow, debt, and subsequent financing.

This in turn explains why Google had to move on equity. Google Cloud's Q1 2026 revenue increased 63% year over year, and backlogged orders doubled from 230 billion to over 460 billion USD compared to the previous quarter. The demand from clients alone on contracts has far exceeded Google’s capacity expansion speed. In other words, even for a cash cow company like Google, AI capital expenditure has grown to a point where equity dilution had to begin.

Berkshire's 10 billion private placement is another detail that needs to be looked at separately in this financing. Buffett has rarely participated in first-level markets over the past 60 years, and even less so in capital expenditure financing for "new economy" companies. This time, buying at fixed prices of $351.81 for Class A and $348.20 for Class C comes closer to a form of identity authentication, effectively stamping “AI computing power as an infrastructure asset class.”

Microsoft's route vs Google's route is diverging

To understand the real significance of this financing, we need to compare the two largest computing power buyers.

Microsoft is following a “self-built plus Neocloud outsourcing” route. Its self-developed chip Maia is progressing slower than expected, while the computing power demand for OpenAI training and inference is growing exponentially. Since the end of 2025, Microsoft’s contractual commitments to the Neocloud system have exceeded 60 billion USD: 23 billion for Nscale (for deploying 200K GB300s), with CoreWeave, Nebius, IREN, Lambda Labs sharing the rest. All these contracts exclusively use Nvidia GPUs. Microsoft has to heavily rely on Neocloud because its self-built capacity cannot keep up with demand, and its own chips cannot compete with Nvidia.

Google is taking a different path. With self-developed TPUs, self-built data centers (not relying on Neocloud), and now also selling TPUs to others, it is trying to capture the Neocloud market via a JV with Blackstone. Google does not need Neocloud; it aims to become Neocloud's competitor.

This divergence is the real strategic pivot of this financing round. The deeper Microsoft binds to Neocloud, the more Google has to break Neocloud. The two companies are choosing different paths because of their underlying assets; Microsoft lacks its own high-end AI chips, while Google has TPUs.

What supports Google's route is the real progress of TPUs. Anthropic has already shifted its training tasks to TPUs on a large scale by 2025, while Meta, SSI, and xAI have been reported to be negotiating TPU orders. Internally, Google asserts that the cost-effectiveness of TPUs in specific inference workflows is 3 to 5 times that of Nvidia GPUs, a figure confirmed by multiple independent analysts.

The asymmetric fate of the trio

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

In terms of short-term cash flow, Google does not pose a threat. CoreWeave's Q1 backlog has reached nearly 100 billion USD, including the 21 billion contract signed with Meta in March and multi-year contracts with Anthropic. Nebius had Q1 revenues of 390 million USD, an increase of 841% year-over-year, with 2026 annual guidance of 3 to 3.4 billion in revenue and an annualized operating rate of 7 to 9 billion, and a signed five-year contract with Meta worth 27 billion USD. IREN holds contracts amounting to 9.7 billion with Microsoft and 5.5 billion with Nvidia. These are all locked-in contracts using Nvidia GPUs that Google’s TPU cannot replace.

What is being broken is the valuation narrative. The logic behind the high valuations of these three companies is built on three premises: overwhelming demand for AI computing power, Nvidia GPUs as the only option, and hyperscalers' self-built capacities not keeping pace with demand. Google’s combination of strategies has progressively undermined these three premises. TPUs are a real alternative, new capacity is arriving, and using JVs to accelerate when self-building falls short.

Yet the situations of the three companies are entirely different.

CoreWeave's high valuation risk has been partially released, but its debt leverage has not been cleared. Its market positioning as "AWS of the GPU era" carries the greatest ambition and highest valuation premium. Nvidia already holds about 11% of CoreWeave’s shares, worth nearly 4.9 billion USD, doubling its stake in January 2026 at 87.20 USD per share. This deep binding prevents CoreWeave from pivoting towards TPUs, as in customers' perception, it is simply an agent of Nvidia GPUs. As long as Google’s strategy convinces the market that TPUs truly become a first-line option, CoreWeave’s valuation premium will shrink.

Nebius is in a middle position. Its technology stack is relatively open (Soperator is open-sourced, similar to CoreWeave’s SUNK route), and while its customer base leans towards Nvidia GPUs, it has more flexibility. Nebius's debts and cash are nearly hedged, and Leopold Aschenbrenner, a former OpenAI researcher, established a position in Nebius at the end of May; he only took action after Google entered the market, betting on whether growth or valuation would outpace the other.

IREN is the most unusual. This company transitioned from being a Bitcoin miner and has the heaviest assets and lowest valuation premium among the trio. The cash flow from its contracts of 9.7 billion with Microsoft and 5.5 billion with Nvidia is enough to support its fundamentals. It carries no pressure from the "high valuation narrative" being broken; in the new landscape, IREN has shifted from "the weakest" to "the most stable," although it is no longer cheap.

The computing power market transitions from undersupply to customer stratification

The second-order implication of this is a structural shift in the computing power market.

The AI computing power market has been a typical seller's market for the past 18 months, with Nvidia determining the supply rhythm and all buyers queuing up. Now, three layers are synchronously occurring.

First, leading model laboratories are starting to diversify their stacks. Anthropic has publicly started using Google TPUs, AWS Trainium, and Nvidia GPUs, and OpenAI has also been reported to be evaluating TPUs. Once multi-stack becomes the standard configuration for top laboratories, the “exclusive Nvidia GPU” label for Neocloud will become a limitation from the customers' perspective.

Second, the hyperscaler routes are diverging. Microsoft (deeply tied to Neocloud), Google (self-built plus selling chips plus entering Neocloud market), and Amazon (primarily self-developing Trainium) are all taking entirely different directions. This divergence directly decides Neocloud's customer structure. Currently, Neocloud's key clients are Microsoft and Meta, with Google being completely absent. If Microsoft reduces outsourcing due to improvements with Maia or adjustments in relationships with OpenAI, Neocloud’s revenue side faces structural risks.

Third, capital cost stratification is emerging. Google is financing through equity, Buffett's endorsement, and operating cash flow, with a funding cost close to zero. CoreWeave's latest loan pricing is SOFR (secured overnight financing rate) + 4.5%. In a capital-intensive business where GPU depreciation cycles only last 5 to 7 years, this funding cost gap can result in a lethal difference due to compounding. Neocloud currently exists because Nvidia GPUs are still in high demand; once GPUs transition from scarce goods to relatively abundant commodities, the player with the lowest capital cost will regain market dominance. This is what Google is betting on.

Next, track three indicators

Returning to the 80 billion in equity financing, the real signal it sends to the market is that Google is treating AI computing power as a market needing to be re-divided. The three companies, CoreWeave, Nebius, and IREN, will still be able to run their short-term contracts for another two to three years, but their high valuations based on “Nvidia uniqueness” have already been pried open by Google’s combination of strategies.

Moving forward, tracking three events will suffice: whether the Google-Blackstone JV can timely activate the 500 megawatts of capacity in 2027, whether the TPU customer list can expand from Anthropic to Meta and xAI, and whether Microsoft will revert to negotiations over TPUs if relationships with OpenAI become strained. If any two of these three events come to fruition, the story of the trio will need to be rewritten.