Google adds two specialized TPU systems for 2027

Google attributes the new configurations partly to external customers seeking larger context windows and specialized security. TPU 8tx uses the same silicon as TPU 8t but pairs it with host processors from Advanced Micro Devices, the CPU and GPU supplier known as AMD, and Google’s low-latency Boardfly network. Google claims up to twofold lower latency than TPU 8t and a near-zero performance penalty for the security protections. The announcement does not give workload-level test details for those claims. Its intended customers include governments and regulated businesses that need to run inference—the execution of already-trained models—alongside sensitive information.

The system also supports private fine-tuning, model distillation and reinforcement learning. Google’s pitch is that customers can run and customize models within one protected system, avoiding transfers of sensitive data between separate clusters.

TPU 8i Max addresses a different constraint. A context window is the information a model can consider in a request; large codebases or hours of video can create demanding memory workloads. Google says combining two TPU 8i chips in one package doubles compute, on-chip SRAM and high-bandwidth-memory bandwidth relative to TPU 8i. It also fits twice the silicon into the existing rack footprint. Those are company specifications, not evidence that every application will run twice as fast.

NVIDIA’s August 2023 H100 explanation illustrates confidential virtual machines and encrypted CPU–GPU transfers. This historical topology describes NVIDIA systems, not TPU 8tx. Image: NVIDIA. Source: NVIDIA Technical Blog.

Security is a system feature with existing rivals

AMD says its fifth-generation EPYC server CPUs provide TPU 8tx’s host infrastructure, handling data preparation, workload coordination and communication with the accelerators. Its announcement describes Secure Encrypted Virtualization technology working with the TPU to extend confidential computing beyond the CPU.

The protections include verifying an accelerator’s identity and security configuration before allowing access to protected memory. AMD says these protections help keep prompts, model weights and intermediate data encrypted in CPU memory and during transfers, with hardware isolation during accelerator execution. That is more precise than treating all processing as simply occurring on encrypted data.

AMD calls 8tx the first cloud TPU system with hardware-enforced confidential computing. Nvidia had already documented confidential-computing support on its H100 GPU in an August 2023 technical explanation, describing an early-access release. It outlined secure boot, a protected connection to the CPU environment and a signed attestation report that users could verify. Google is adding the capability to its TPU offering within an established category of accelerator security.

Nvidia’s account also shows why security overhead needs workload testing. It said raw GPU compute and memory bandwidth were comparable with non-confidential operation, but encryption and transfers across unprotected interconnects could limit throughput when workloads performed little computation per byte of input. Those historical H100 observations are not benchmarks for TPU 8tx.

AMD is also a supplier across competing architectures. It says EPYC processors host systems using Nvidia GPUs, AMD’s own Instinct GPUs and custom accelerators. Its role in 8tx extends an existing Google Cloud collaboration across processor generations and security, including Google reviews of AMD security-processor firmware. AMD cautions that security features vary by processor generation and must be enabled by the server manufacturer or cloud provider.

The economic test goes beyond chip specifications

Google’s internal AI development gives its chip teams early access to Gemini models for finding bugs and improving power efficiency, according to the report. External adoption also depends on software. Google says its recently open-sourced TorchTPU lets code written in the PyTorch framework and models from Hugging Face run natively on TPUs. It also lists TPU support for the open-source inference engines vLLM and SGLang. Its Agent Compute Engine software suspends idle agent sessions and moves working memory off the accelerator, targeting the cost of keeping expensive hardware occupied while agents wait for tools or human input.

Other software changes address memory movement and distributed execution. Google says its open-source TPU Sync coordinates model state across accelerator memory, host memory, storage and networking. Google is making its Cloud Pathways distributed runtime and dynamic allocation of TPU groups generally available on Google Kubernetes Engine. These changes address utilization alongside the new hardware.

The commercial opportunity is substantial, but forecasts should not be mistaken for booked revenue. The report cites a Bloomberg Intelligence projection that TPU sales, including related hardware and software, will exceed $135 billion in 2028, compared with an estimated $12 billion in 2026. Those figures exclude customers’ TPU usage through Google Cloud. It also cites a regulatory filing giving Anthropic, the maker of the Claude chatbot, an option to buy as much as $150 billion in chips between 2028 and 2031. An option is not a committed purchase.

Customer experience offers a narrower measure of competitive progress. Jump Trading, a trading firm using AI for work including trading research, tested TPUs for several years before signing a larger Google contract about a year ago. Its chief technology officer, Alex Davies, told the report it has thousands of TPUs but still mainly uses Nvidia for training. He described TPUs as effective for a subset of tasks, with better power efficiency where they matched or exceeded Nvidia’s performance.

Mandeep Singh, an analyst at Bloomberg Intelligence, estimates in the report that TPUs cost about 30% less than Nvidia GPUs while providing competitive throughput. The account does not supply the chip configurations, workload or contractual terms needed to turn that estimate into a purchasing comparison for 8tx or 8i Max. Nvidia is also moving toward more specialized systems, including inference-focused hardware using technology acquired from Groq and CPUs for managing agents.

Prospective rivals include Tenspresso, a startup co-founded by two former TPU engineers targeting long-context inference. Its first chip was at least a year from design completion, the report said. Anthropic and prospective TPU buyer OpenAI are also pursuing in-house chips staffed with TPU veterans.

Singh also says Google’s TPU margins are lower than Nvidia’s. The report flags a possible slowdown in AI infrastructure spending. Rapid sales growth, profitable growth and sustained demand are separate tests of the business.

What customers still need before 2027

The immediate test is delivery. The report relays Google’s earnings statement that every available TPU was already contracted and describes shortages of memory and manufacturing capacity. Google originally developed TPUs alongside Broadcom, which helped with design and production coordination. The report describes a deal with Broadcom rival Marvell that includes TPU work, while noting limited details and Vahdat’s refusal to comment on Marvell’s role.

Oracle said it would offer TPUs to its cloud customers starting in 2027, adding to their existing Nvidia and AMD choices. Buyers still need configuration-specific pricing, delivery commitments and workload tests that include security overhead, memory behavior and software migration. The announcements do not provide those answers for the two new systems.

Google has established what it wants each configuration to do. The unresolved question is how much useful work customers can obtain, at what total cost, from systems they can actually secure in 2027.

Meta expands Instagram’s AI age checks globally without publishing error rates

Meta is expanding an existing system that places suspected teenagers into Instagram’s protective settings despite adult birthdays. Its global rollout follows legal settlements, but the announcement provides no numerical accuracy results.