Google unveiled two AI chips at Cloud Next 2026 on 22 April delivering 2.8× the training performance of its prior generation — and OpenAI is now buying Google TPU capacity.
A single chip with 384 megabytes of on-board SRAM — triple the memory of its predecessor — is Google's latest bid to unseat Nvidia as the default processor of the global AI industry.
Google unveiled two new tensor processing units at its Cloud Next conference in Las Vegas on 22 April 2026: the TPU 8t, designed for training large AI models, and the TPU 8i, optimised for inference tasks. Together, they represent the company's most targeted competitive challenge to Nvidia's H100 and B200 lines. The announcement carried an unusual commercial signal: OpenAI, which has trained almost exclusively on Nvidia hardware since its founding, is now purchasing TPU capacity from Google Cloud.
Google Cloud · TPU 8 · artificial intelligence chips
The performance claims are substantial. Google says the TPU 8t delivers 2.8 times the training throughput of its seventh-generation Ironwood chip, announced in November 2025, at the same price per unit. The TPU 8i improves inference performance by 80% over Ironwood. Both chips contain 384 MB of SRAM, compared with Ironwood's 128 MB — a tripling of on-chip memory that reduces the time models spend waiting on external memory access, one of the primary latency bottlenecks in large-scale inference. Commercial availability is set for "later in 2026," Google Cloud said at the conference, without specifying a quarter.
Continue reading to see the full article
“Google says the TPU 8t delivers 2.8 times the training throughput of its seventh-generation Ironwood chip, announced in November 2025, at the same price per unit.”
Anthropic, the San Francisco-based AI safety company and maker of the Claude model family, simultaneously committed to purchasing what it described as "multiple gigawatts" of Google TPU capacity — a figure that, if realised, would place it among the world's largest compute buyers. The Anthropic announcement reinforced the companies' existing partnership and signaled that frontier AI demand for compute remains far ahead of current supply from any single vendor.
Key Takeaways
→Google Cloud: Google says the TPU 8t delivers 2.
→TPU 8: Google says the TPU 8t delivers 2.
→artificial intelligence chips: Google says the TPU 8t delivers 2.
→Nvidia rivalry: Google says the TPU 8t delivers 2.
For Google, the TPU 8 generation is a chance to convert the AI industry's insatiable appetite for compute into Cloud revenue. Google Cloud's revenue grew 28% year-over-year in Q1 2026, reaching $12.4 billion, according to Alphabet's April 2026 earnings report — but its share of cloud infrastructure market remains below both Amazon Web Services and Microsoft Azure. A credible TPU that attracts frontier labs away from Nvidia accelerators is a structural shift, not a product cycle.
Nvidia's position is not yet threatened at scale. The company shipped approximately $40 billion in data center GPUs in the fiscal year ending January 2026, according to its own reporting, and its CUDA software ecosystem gives it a switching-cost advantage no competitor has overcome. AMD's MI350 chips, launched in March 2026, have attracted modest enterprise interest but have not dented Nvidia's share of frontier AI training. Google's own previous TPU generations found significant adoption internally but struggled to attract third-party workloads at scale — the gap between benchmark performance and real-world migration has been wider than Google's roadmaps implied.
Advertisement
Google Cloud · TPU 8 · artificial intelligence chips
The caveat buried in Google's announcement is software. Training on TPUs requires rewriting or recompiling workflows designed for Nvidia GPUs — a process major labs have estimated at three to six months of engineering time per model architecture. Google has improved its JAX and XLA compiler tooling significantly since 2024, and Anthropic's commitment suggests those improvements are paying off. But for labs that have not made the transition, switching cost remains real, and "later in 2026" availability means the competitive test is still months away. The question of whether OpenAI's TPU purchase is a pilot or a structural shift in its infrastructure strategy is one the company has not answered publicly.
The next inflection point is Nvidia's GTC conference, tentatively scheduled for September 2026, where the company is expected to reveal pricing and availability for the Blackwell Ultra B300 series. If Google can sign additional frontier labs — particularly xAI or Meta's research division — before that announcement, it will have materially narrowed the gap Nvidia has spent a decade building.
How much faster is the Google TPU 8t than its predecessor?
Google says the TPU 8t delivers 2.8 times the training performance of the seventh-generation Ironwood TPU, announced in November 2025, at the same cost. Both new chips contain 384 MB of SRAM — triple Ironwood's 128 MB — which reduces latency from external memory access.
Is OpenAI using Google TPUs instead of Nvidia GPUs?
As of April 2026, OpenAI is purchasing Google TPU capacity alongside its continued use of Nvidia hardware. OpenAI has historically trained its models almost exclusively on Nvidia GPUs. The Cloud Next announcement is the first confirmed public signal that OpenAI is diversifying its compute suppliers.
When will Google's TPU 8t and 8i be commercially available?
Google Cloud said "later in 2026" at its Cloud Next conference on 22 April 2026 but did not specify a quarter. The chips are not yet generally available as of April 26, 2026.
How do Google TPUs compare to Nvidia GPUs for AI workloads?
TPUs are purpose-built for matrix operations used in deep learning, while Nvidia GPUs are more general-purpose accelerators backed by the CUDA software ecosystem. Google claims TPU 8 outperforms comparable Nvidia hardware on training and inference benchmarks, but most AI labs still build on Nvidia due to software compatibility and established tooling.
