SemiAnalysis said in a report published Aug. 7 that Google’s Gemini lineup may have already lost momentum in the frontier AI model race, while Google Cloud Platform, or GCP, is building a larger business by selling TPU capacity and AI infrastructure to outside customers.
In SemiAnalysis’ view, that trade-off may still work for Alphabet shareholders. The firm estimated that as Google sells more TPUs to external customers such as Anthropic, GCP’s revenue growth in 2027 could exceed 100%, well above current market expectations. The argument running through the report is that Google may be moving away from trying to build the world’s strongest model itself and toward supplying the compute used by whoever does.
Leadership changes at DeepMind shaped the report’s bearish Gemini call
SemiAnalysis began with the changes inside Google DeepMind. Google announced on Aug. 5 that DeepMind’s leadership structure had been reshuffled. DeepMind co-founder Demis Hassabis will no longer handle day-to-day operations, while Jeff Dean, the longtime Google engineer, Google Brain co-founder, and a central figure behind the TPU program, has left Google to launch a new lab called Discovery Loop.
Dean was not the only departure cited in the report. SemiAnalysis also pointed to the exits of Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, all described as heavyweight researchers. Vinyals had been a co-lead on Gemini. Former DeepMind CTO Koray Kavukcuoglu has now taken over DeepMind and Gemini.
From that, SemiAnalysis drew a stark conclusion: DeepMind is no longer functioning as a truly competitive frontier lab, and Gemini 3 Pro may have been the high-water mark for Google in this cycle of frontier model competition.
The report said Gemini 3 Pro, when launched in November 2025, was briefly seen as one of the strongest models in the world and even pushed OpenAI into a “Code Red” posture. But SemiAnalysis argued that Google then started falling further behind OpenAI and Anthropic. It said Gemini 3.5 Flash underperformed expectations, Gemini 3.5 Pro was canceled, Gemini 3.6 Flash was used as a bridge, and Google has now pinned its hopes on Gemini 4.
API usage is still rising, but growth has slowed
SemiAnalysis also used token consumption data to support its weaker view on Gemini. According to the report, Gemini’s first-party API grew from 10 billion tokens per minute to 16 billion tokens per minute in the first quarter of 2026, a quarter-on-quarter increase of about 60%.
In the second quarter, that figure climbed again to 22 billion tokens per minute, but the quarterly growth rate slowed to about 38%. SemiAnalysis said the deceleration in token growth is now beginning to show up in revenue from Gemini’s first-party API.
That slowdown, however, does not mean Google’s enterprise AI business is stalling at the same pace. Google’s official second-quarter results showed that its model APIs processed roughly 22 billion tokens per minute, Gemini Enterprise had reached nearly 90% of Fortune 100 companies, Google Cloud revenue rose 82% year over year to $24.8 billion, and Cloud backlog reached $514 billion.
One reason, the report said, is that Google Cloud is not limited to selling Gemini. Enterprises can use third-party models, including Claude, through Google’s AI platform. That means even if Anthropic beats Gemini at the model level, Google can still make money from Anthropic and from enterprise customers using Claude.
Google is selling TPU capacity to Gemini’s rivals
SemiAnalysis said the deeper issue inside Google’s AI strategy is not the model itself but compute allocation: whether scarce compute should stay with DeepMind or be sold into the market.
For any frontier AI lab, compute is a core production input. OpenAI and Anthropic have both been trying to secure as much GPU and data center capacity as possible. Google, by contrast, has chosen to commit large amounts of in-house TPU capacity to external customers, including Anthropic, which the report described as Gemini’s most direct competitor.
SemiAnalysis estimated that more than 20% of TPU shipments from the third quarter of 2026 through the fourth quarter of 2027 will be sold directly to Anthropic. That estimate does not include the hundreds of thousands of TPUs GCP has already leased to Anthropic, or additional capacity promised over the coming quarters to Anthropic and Meta.
The report framed this as an unusual commercial loop: Google develops TPUs, GCP sells TPU and data center compute to Anthropic, Anthropic uses that compute to train Claude, and Claude then competes with Gemini.
From DeepMind’s perspective, that can look like funding a rival. From the perspective of Google Cloud CEO Thomas Kurian, SemiAnalysis said, it is a rational infrastructure business. Kurian has long positioned Google as a platform company and has argued that TPUs should not exist only as Gemini’s private AI chips but as general infrastructure for enterprise workloads, high-performance computing, and outside AI labs.
SemiAnalysis said the internal contest between “DeepMind wants compute” and “GCP wants to sell compute” now appears to be breaking in Kurian’s favor.
Google Cloud’s business model is shifting beyond leasing compute
The report said Google is no longer just renting out TPU capacity. Alphabet confirmed in its second-quarter results that Google Cloud began directly delivering TPU systems into customer data centers for the first time in the second quarter, and the company has already started recognizing that revenue.
Google also said Cloud revenue growth remained materially stronger even after excluding TPU system sales. SemiAnalysis estimated TPU system sales contributed about $1.2 billion in the quarter. Excluding that amount, it said the core GCP business still posted year-over-year growth a little above 70%.
The larger number, in SemiAnalysis’ telling, is future demand. The firm estimated that Google’s TPU system backlog now exceeds $150 billion. As those orders are recognized over time, GCP’s annual growth rate in 2027 could reach the “mid-100%” range, clearly ahead of the roughly 64% currently expected on Wall Street.
Google did not break out a TPU-specific backlog figure, but it did confirm that total Google Cloud backlog hit $514 billion in the second quarter, up by more than $50 billion from the prior quarter. The company expects a little more than half of that amount to turn into revenue over the next 24 months.
SemiAnalysis sees Google’s external AI business nearing $200 billion by late 2027
The report’s forecast becomes more aggressive when it turns to the size of Google’s outside AI revenue opportunity. SemiAnalysis estimated Gemini ARR was about $12 billion in the second quarter of 2026. By comparison, it said third-party AI IaaS/TaaS ARR through GCP alone could exceed $73 billion by the end of 2027, with TPU system sales adding roughly another $120 billion.
Combined, that would put Google’s external AI business close to $200 billion.
SemiAnalysis also estimated that even if direct TPU system sales carry lower margins than traditional cloud services, Google Cloud’s overall EBIT margin could still stay in the high-30% range and contribute about $3 in additional EPS to Alphabet in 2027.
The real question is model leadership versus capital allocation
SemiAnalysis said the point of the report is not simply whether Gemini can beat GPT or Claude. The deeper issue is how Google allocates capital and compute.
If Google keeps a large share of TPU capacity for DeepMind, it would be giving up near-term cloud opportunities worth tens of billions, or even more than $100 billion, in exchange for a chance that Gemini could regain the lead. If it sells that capacity to Anthropic, Meta, and other AI companies, GCP could quickly become one of the largest compute suppliers of the AI era. It would also mean handing Google’s most valuable computing resources to Gemini’s competitors.
SemiAnalysis’ position is that Google has already made that choice. Gemini may lose the frontier model war, but Google may still avoid losing the larger AI war.