SemiAnalysis said on 15 October 2026 that Google has submitted its first TPUv7 results to SemiAnalysis’s Agentic Inference Benchmark, running at FP8 on the vLLM inference engine, and that they showed 1.67x better performance per dollar than Nvidia’s GB300 NVL72 running at FP4.
The figure comes from a research firm’s benchmark, not from Google or Nvidia, and it is a preview: SemiAnalysis’s InferenceX dashboard labels the TPUv7 line “Official Preview: Results may change as validation and publication continue.”
Business Pill · TOTAL COST OF OWNERSHIP
A one-minute explainer of total cost of ownership: why the cheapest price is not always the cheapest option. It teaches the general idea only and says nothing about any organisation in this story.
The key insight: As we read it, the result is as much about price as about throughput: SemiAnalysis’s own cost model puts a TPUv7 chip-hour at $1.03 against $2.31 for GB300, so the per-dollar ranking rests on those cost assumptions as well as on tokens served.
The test is SemiAnalysis’s AgentX scenario on Qwen3.5 397B. SemiAnalysis says AgentX captures four elements of an agentic workload: many user and assistant turns per session, long context, high prefix reuse and bursts of short-lived sub-agents.
“TPU has been able to demonstrate that its upcoming TorchTPU vLLM stack was able to successfully handle agentic workloads,” SemiAnalysis wrote.
The dashboard lists 872,048 completed requests across its plotted points, with a median input sequence length of 83k tokens.
SemiAnalysis says Google’s first TPUv7 submission to its agentic inference benchmark (FP8, vLLM, Qwen3.5 397B) showed 1.67x better performance per dollar than Nvidia’s GB300 NVL72 at FP4. SemiAnalysis labels the result a preview that may change.
What Was Posted on X
SHOCKING: We have confirmed that multiple Neoclouds have been refused access to GTC Berlin due to them adding AMD GPUs, TPUs, and other chips to their cloud offerings
Note that this doesn’t apply to hyperscalers, as they have more buying power. But for neoclouds, executives feel NVIDIA has been using high-pressure tactics to keep them NVIDIA-only.
As such, neocloud executives are starting to consider offering TPUs or AMD GPUs quietly to avoid getting pressured by NVIDIA and to avoid the feeling that NVIDIA will retaliate. https://x.com/SemiAnalysis_/status/2071005685903454496
SemiAnalysis said on 16 October that it had confirmed several neoclouds were refused access to Nvidia’s GTC Berlin after adding AMD GPUs, TPUs and other chips to their clouds, that hyperscalers were not affected, and that some neocloud executives are considering offering TPUs or AMD GPUs quietly. These are SemiAnalysis’s claims; we found no Nvidia statement on them. The post links an earlier SemiAnalysis post (x.com/SemiAnalysis_/status/2071005685903454496).
BREAKING NEWS๐จ: @Inferact & Google submit their FIRST preliminary TPUv7 Kimi K3 2.8T Model Submission, and it is already delivering strong perf per TCO against GB200 NVL72, B300, and MI355 in the ultra-high-interactivity regime due to their megakernel approach for external customers.
One of the most important aspects of an external compute platform is a strong community, and what makes this submission special is that Inferact (Makers of vLLM) led the submission end to end using the TorchTPU vLLM stack.๐๏ธ(1/6)๐งต
SemiAnalysis’s 18 October thread opener on the first preliminary TPUv7 Kimi K3 2.8T submission, which it says Inferact led using the TorchTPU vLLM stack. The attached chart plots total tokens per $1 of TCO against P90 interactivity in its ‘Owning at Large Hyperscaler Volume’ tier; the comparison and the cost rates are SemiAnalysis’s, and the TPU line is labelled a preview that may change.
#Googleโs TPU v9, involving #MediaTek, could face delays, with shipments potentially slipping to 1Q28. Meanwhile, #ByteDance may purchase data center chips from MediaTek. What other Taiwanese firms are riding the AI #ASIC wave?๐กMore: https://pse.is/9rrmlz ๐
TrendForce’s own post on 19 October summarising the Commercial Times report it relays: Google’s TPU v9, involving MediaTek, could face delays, with shipments potentially slipping to the first quarter of 2028. The timing comes from institutional investors cited by Commercial Times, not from Google or MediaTek, and the post keeps that hedge, so we treat the slip as reported, not confirmed.
How Wide the Gap Is
“What is surprising is that Google TorchTPU vLLM on TPUv7 FP8 with just 4 chips was able to beat GB300 NVL72 FP4 on performance per TCO across most of the pareto frontier,” SemiAnalysis wrote.
The same post says GB300 at FP4 still wins at low interactivity, where TPUv7 at FP8 still comes out ahead of GB300 at FP8.
SemiAnalysis’s chart plots total tokens per $1 of TCO against P90 interactivity, the speed each user sees in tokens per second. It marks gaps of 1.67x, 1.46x and 1.92x along the TPUv7 curve at different speeds, and SemiAnalysis’s post names GB300 NVL72 FP4 as the comparison for the 1.67x.
The Cost Basis Behind the Number
The dashboard prices each chip from SemiAnalysis’s July 2026 pricing surveys and AI cloud TCO model, at a cost tier it calls owning at large hyperscaler volume: $1.03 per chip-hour for TPUv7 and $2.31 for GB300.
By our arithmetic, that puts a GB300 at about 2.2 times the hourly cost of a TPUv7 in SemiAnalysis’s model, so the per-dollar result depends on those cost assumptions as well as on throughput.
The chart labels the GB300 runs Dynamo SGLang and the TPUv7 run vLLM. The dashboard adds a note on how it counts chips: “Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate cost per million tokens per decode chip or per prefill chip, rather than per total chip count. This makes direct token cost comparison with aggregated configs not an apples-to-apples comparison.”
SemiAnalysis’s Other Result This Week
On 14 October, about four hours before the TPU thread, SemiAnalysis published a separate result for Nvidia’s own stack. It said vLLM, Inferact and Nvidia used an optimisation called wide expert parallelism, or wideEP, on GB300 NVL72 to reach 7.2x better performance per TCO than every other AI chip.
That result is on a different model, DeepSeek v4.1 Flash. “Currently, Inferact & NVIDIA have only submitted up to 200 tok/s/user p90 interactivity SLO,” SemiAnalysis wrote.
As we read it, the two results come from different models and configurations, so they do not rank the same chips on the same test.
Who Is Behind the Submission
“Great work by the Google TPU, TorchTPU, @Inferact, @RedHat, and vLLM teams on their first amazing AgentX result,” SemiAnalysis wrote, adding that a full article with many more models is still to come.
SemiAnalysis also wrote that it is “amazing to see Google go full steam ahead with externalizing the power of TPUs outside of Google.”
When we searched X on 15 October, we found no post by Google, Google Cloud or the vLLM project about the result.
We also reported on Gemini 4 Argon‘s paid rollout at $2 and $10 per million tokens, and on Gartner saying more than $3.5 trillion of AI infrastructure spending could face impairment by 2030.
Update, 19 October: A Second TPUv7 Preview, This Time on Kimi K3
SemiAnalysis posted on X on 18 October 2026 that Inferact, which it describes as the makers of vLLM, and Google had made their “FIRST preliminary TPUv7 Kimi K3 2.8T Model Submission” to its InferenceX benchmark, and that it “is already delivering strong perf per TCO against GB200 NVL72, B300, and MI355 in the ultra-high-interactivity regime”. SemiAnalysis said Inferact “led the submission end to end using the TorchTPU vLLM stack”.
SemiAnalysis kept two cost bases apart. “On an internal Google cost basis, this initial preliminary TPUv7 Kimi K3 is already competitive against even GB300 NVL72”, it wrote, adding that Nvidia has already submitted multiple updates for GB300. For external customers it stated a belief rather than a result: with other combinations of prefill-decode disaggregation ratios in upcoming submissions, “we believe it will be competitive against even GB300 NVL72 on an external customer cost basis”. It called this “a preliminary submission, with multiple follow-up submissions ahead”.
The chart attached to the thread, in SemiAnalysis’s “Owning at Large Hyperscaler Volume” cost tier, lists TPUv7 at $1.21 per chip-hour against $2.31 for GB300, $1.86 for GB200, $2.26 for B300 and $1.50 for MI355X, taken from its July 2026 pricing surveys and AI cloud cost model, and labels the TPU line an official preview whose results may change. When we read the Kimi K3 dashboard page on 19 October it listed TPUv7 at $1.03 per chip-hour in the same tier, the rate we cited on 15 October; the SemiAnalysis texts we read do not say why the two figures differ.
As we read it, the second preview widens the test to a 2.8-trillion-parameter model, and SemiAnalysis’s own split between an internal Google cost basis and an external customer one marks the gap that matters for anyone renting TPUs rather than owning them.
Update, 19 October: Taiwan Supply-Chain Reports Say TPU v9 Could Slip to 2028
TrendForce reported on 19 October that, according to Taiwan’s Commercial Times, citing institutional investors, “Google’s TPU v9 project, involving MediaTek, could see its tape-out delayed until January 2027 following design revisions, with shipments potentially pushed back to the first quarter of 2028.”
The report says “revisions to Google’s TPU v9 Core A design, codenamed Humufish, have delayed development by approximately three weeks, affecting the Core B and I/O designs handled by MediaTek.” It adds: “Although the November 2026 sampling target remains unchanged, the initial tape-out is now expected in January 2027, followed by the production-version tape-out in the second quarter.” A GF Securities note summarised by Zhitong Finance, whose body we could not read, is headlined as saying TPU v9 progress is “slightly delayed”.
Neither the TrendForce report nor the Commercial Times article it cites carries a comment from Google or MediaTek. The same Commercial Times report, citing institutional investors, projects global AI ASIC accelerator shipments of 14.1 million units in 2027, up about 69% year over year.
As we read it, if the reported slip holds, samples could still arrive on the November 2026 target while customer shipments wait until the first quarter of 2028.
The Structural Read
As we read it, the comparison SemiAnalysis draws is per dollar of total cost of ownership rather than per chip: its chart plots total tokens per $1 of TCO, which is how a four-chip TPUv7 run can be set against a GB300 NVL72 system.
As we read it, the caveats matter as much as the headline: SemiAnalysis says GB300 at FP4 still wins at low interactivity, and its own dashboard notes that disaggregated and aggregated configurations are not an apples-to-apples token-cost comparison.
As we read it, the other signal is who submitted: SemiAnalysis credits the Google TPU, TorchTPU, Inferact, Red Hat and vLLM teams, and describes Google as externalizing the power of TPUs outside of Google.
“What is surprising is that Google TorchTPU vLLM on TPUv7 FP8 with just 4 chips was able to beat GB300 NVL72 FP4 on performance per TCO across most of the pareto frontier”
Three Implications
GOOGLE SemiAnalysis says this is Google’s first TPUv7 submission to its agentic benchmark, and that it expects Google to submit many more models soon.
NVIDIA SemiAnalysis’s separate 14 October result puts GB300 NVL72 with wideEP at 7.2x better performance per TCO than every other AI chip, on a different model, DeepSeek v4.1 Flash.
BENCHMARK READERS SemiAnalysis labels the TPUv7 numbers an official preview that may change as validation and publication continue, with its full article still to come.
The framework puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, a per-dollar benchmark like SemiAnalysis’s makes that marginal cost visible: on this test, the chip-hour price in SemiAnalysis’s model, $1.03 for TPUv7 and $2.31 for GB300, sits inside every token served.
SemiAnalysis’s TCO assumption per chip-hour by chip, from its InferenceX dashboard (cost tier: owning at large hyperscaler volume): $1.03 for TPUv7 and $2.31 for GB300.
What Is Not Established
SemiAnalysis labels the TPUv7 numbers a preview that may change, and it has not yet published the full article it promises, so the methodology behind each point is not established here.
The result covers one model, Qwen3.5 397B, and SemiAnalysis’s own cost assumptions per chip-hour, not prices Google or Nvidia charge. The TPUv7 runs at FP8 and the GB300 comparison at FP4, as the chart labels them.
Business Engineer Framework
The State of The Inference Economy
A Business Engineer framework on why inference, not training, now sets who earns what in AI, and where the cost per token is decided.
SemiAnalysis said on 15 October 2026 that Google’s first TPUv7 submission to its agentic inference benchmark showed 1.67x better performance per dollar than Nvidia’s GB300 NVL72 at FP4, on Qwen3.5 397B.
It calls the numbers a preview, says GB300 at FP4 still wins at low interactivity, and its cost model puts a TPUv7 chip-hour at $1.03 against $2.31 for GB300.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read SemiAnalysis’s own X thread of 15 October 2026 on Google TPUv7 in full, its InferenceX Qwen3.5 dashboard page and the two charts attached to the thread, and its separate 14 October thread on GB300 wideEP. The figures and quotations come from those SemiAnalysis texts; the result is a preview SemiAnalysis says may change, and SemiAnalysis’s full article had not been published when we wrote. The TSMC, Nvidia, Gemini and Gartner details come from our own earlier reporting, and the 2.2 times ratio is our arithmetic. The 19 October update draws on SemiAnalysis’s 18 October X thread (posts 1 to 6), the chart attached to it and its Kimi K3 InferenceX dashboard page, all read in full on 19 October; we did not find a separate Google or Inferact post on this submission when we checked their X accounts. The second 19 October update draws on TrendForce’s 19 October report and the Commercial Times article it cites (both read in full; the Commercial Times article is in Chinese and our quotes are TrendForce’s English wording), and on the headline of a GF Securities note summarised by Zhitong Finance on Webull, whose body was not readable to us. These are sourced reports about an unannounced chip; Google has not confirmed them in any of these sources.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
