Picture a hospital that wants an AI assistant to read thousands of patient scans. The technology exists. The doctors are ready. But the data can never leave the building.
Patient records are private by law, and the moment they travel to a cloud server, someone has to trust that server completely. For years, that trust problem has quietly limited what AI can do in hospitals, banks and government offices.
On Monday, Google announced two new chips designed to attack exactly that wall. The TPU 8tx and the TPU 8i Max are the newest members of its eighth-generation Tensor Processing Unit family.
Together they show how the AI chip race is changing: no longer just about raw speed, but about secrecy, memory, and who controls the machines behind the world’s biggest models.
Google TPU 8tx: AI That Keeps Secrets While It Works
The first of the two chips is the TPU 8tx, and it solves a problem most people never think about. Normally, data has to be decrypted before a chip can work on it.
It is protected while it sits in storage, and protected while it travels across a network. But at the moment of use, it is exposed. That brief moment is the gap confidential computing is designed to close.
The TPU 8tx builds security directly into the hardware, keeping sensitive data and model weights isolated even while they are being actively processed. Google is aiming the chip squarely at regulated enterprises and governments.
Think of a bank fine-tuning a model on customer records, or a health service training an assistant on patient data. With the 8tx, Google says, that work can happen without the data ever leaving a protected environment.
The chip is built on the same silicon as the TPU 8t, Google’s training chip from earlier this year. What is new is everything around it. The 8tx is paired with fifth-generation AMD EPYC processors as host CPUs.
That is the first time AMD silicon has ever sat inside a TPU system. Google says the setup delivers up to twice the latency reduction of the TPU 8t, with a near-zero penalty in performance for the extra security.
One system can fine-tune models on private data, shrink large models, and train AI agents, all without moving sensitive data between clusters.
TPU 8i Max: Built to Break the Memory Wall
The second chip attacks a completely different bottleneck. The TPU 8i Max is not about secrecy. It is about memory.
Modern AI agents no longer answer a single question and stop. They work across entire codebases, sift through hours of video, and hold long conversations that stretch over thousands of pages.
The amount of information an AI system can consider in one request is called its context window. As those windows grow, the chips running them hit what engineers call the memory wall.
The processor is fast enough, but it keeps waiting for data. The TPU 8i Max answers that by putting two TPU 8i chips into a single package.
According to Google, the result is double the compute, double the on-chip SRAM memory and double the high-bandwidth memory speed of a single 8i, all inside the same rack footprint.
That suits coding assistants, mathematical reasoning and drug discovery, where the model must hold enormous amounts of information at once.
Four New TPUs in a Single Year
Step back, and the bigger story is the pace. Google introduced the TPU 8t for large-scale training and the TPU 8i for real-time reasoning at its Google Cloud Next event in April.
With the 8tx and 8i Max, that makes four new TPU systems in a single year, something Google says it has never done before. Neither of the new systems is a brand-new processor.
Both are built from existing eighth-generation silicon, repackaged with different memory, networking and host processors around it. Google describes the approach as deliberate, a way to deliver new hardware variants in months rather than years.
Google Cloud revenue more than doubled to $30.71 billion in that quarter. Alphabet shares rose more than 3% in early trading on Monday after the announcement.
Why This Matters in the Fight Against Nvidia
For years, Nvidia has dominated the market for AI chips, and almost every major AI model has been trained on its hardware. Google’s TPUs were originally built for internal use, powering Search, YouTube and Gemini.
But the company has steadily been making them available to outside customers through its cloud business. According to Google, TPUs now serve nearly 100 customers, including frontier AI labs, trading firms, retailers and universities.
Google did not name them, but the direction is clear: it wants to be a serious alternative to Nvidia, not just for its own products, but for the whole AI industry.
The two new chips sharpen that challenge in different ways. The 8tx gives regulated customers a confidential-computing option for their most sensitive work. The 8i Max gives developers the memory headroom that long-context agents demand.
The Models Running on These Chips Keep Getting Bigger
The timing of the announcement is no accident. The models that would run on these chips are growing more capable by the month. Anthropic’s newest model, Claude Fable 5.2, can coordinate more than a thousand AI agents to build a working operating system from scratch, according to the company.
Those experiments need enormous amounts of compute, and they need it in forms that fit the work. A coding agent that reads a whole codebase needs memory. A hospital model that reads patient records needs secrecy.
The TPU 8tx and 8i Max are Google’s answer to both demands at once. The same logic applies beyond chatbots. Nvidia’s new GR00T 2 robot foundation model learns physical skills by watching human video.
That task needs huge training runs and fast inference working together. The infrastructure underneath AI is becoming as important as the models themselves.
What Google Did Not Say
For all the detail in the announcement, some important questions went unanswered. Google gave no availability date and no price for either system, and no eighth-generation TPU has a published price at all.
Amin Vahdat, Google’s chief technologist for AI and infrastructure, told Bloomberg that both systems are expected to launch next year, in 2027. Oracle’s cloud is also expected to start offering TPUs the same year.
There is another caveat worth noting. Every performance claim so far comes from Google itself. No independent test of either system has been published.
The security design of the 8tx has also been described in fragments rather than in a full technical document. Customers in regulated industries will want to see the details before they trust their most sensitive data to it.
What Happens Next
The real test comes in 2027, when both chips should reach customers. If the 8tx delivers its promised security without slowing workloads, banks, hospitals and governments could finally bring their most sensitive data to modern AI.
The contest is no longer only about who builds the fastest single chip. It is about who can build the right chip for each job, and deliver it fast enough to matter.
Google has placed its bet: four specialised systems in one year, each aimed at a different bottleneck. Nvidia, which still leads by a wide margin, will not stand still.
The next round of the race has already begun, and the winners will be decided not in benchmarks, but in data centres.
