The Private BankerSunday, November 22, 2026

Google Splits Its TPU Line Into Specialized 8tx and 8i Max Variants, Pressing Nvidia as Alphabet Gains Premarket

Alphabet unveiled two reconfigured AI chips due next year, one for encrypted workloads paired with an AMD processor and one for long-context inference, as shares rose up to 3.2% in premarket trading.

Close-up of a microchip on a green circuit board
Photo by Waterhouse on Pexels.com

Alphabet’s Google introduced two specialized variants of its Tensor Processing Units on Monday, the TPU 8tx and the 8i Max, widening its challenge to Nvidia in AI accelerators. Alphabet shares rose as much as 3.2% in premarket trading in New York.

The 8tx is built for confidential workloads of the kind run by government agencies and regulated industries. It pairs with a processor from Advanced Micro Devices and encrypts data at the chip level, closing a gap in which information normally has to be unlocked in memory while it is processed. The 8i Max places two TPU chips in a single package to add compute and memory for applications that need long context windows, including drug discovery, mathematics and coding. Both are due to be available next year.

Neither is a new silicon design. Amin Vahdat, Google’s chip chief, said the parts are reconfigured versions of existing designs, adapted to tasks customers have been requesting. “We could build a single chip that tries to address the whole market, or by building specialized versions of the chips, we actually can deliver much more value,” he said. Google had previously released a new AI processor roughly every two years; this year it has put out two at once, bringing its current TPU lineup to four offerings tuned to different needs. No clock speeds, memory capacities or performance figures were disclosed for either variant.

The launch lands as Nvidia shares are already under pressure from a board investigation into Chief Executive Jensen Huang’s ties to a woman suspected of working for Chinese intelligence, and after a sharp selloff in AI-linked stocks last week. Nvidia has responded to the shift toward inference work with CPUs and inference-focused chips that use technology it acquired from the startup Groq.

TPU customers already include Apple, Meta Platforms and Oracle. Oracle said last month it will let its cloud customers use TPUs starting next year; they previously chose between Nvidia and AMD chips. Anthropic holds an option to buy up to $150 billion of chips between 2028 and 2031, a recent regulatory filing shows, and OpenAI is planning TPU purchases. Jump Trading disclosed last month that it uses thousands of TPUs to train models for trading research.

The market for the chips is far larger than a year ago. One projection puts TPU sales, including related hardware and software and excluding capacity rented through Google Cloud, above $135 billion by 2028, up from an estimated $12 billion this year. Google has said every TPU it has available is already contracted to a customer.

Several points cut against a clean share shift. TPUs cost about 30% less than Nvidia’s GPUs while delivering competitive throughput, but Google’s profit margins on them are thinner than Nvidia’s. Google is also one of Nvidia’s largest customers and remains heavily dependent on its GPUs. Jump Trading still uses Nvidia mainly for training, and only a vocal minority of its staff prefer TPUs. Demand for Nvidia’s processors also continues to exceed what the company can supply, which leaves room for rival parts. A backlash against data centers after Democratic gains in the midterm elections could trim planned technology spending of $1.4 trillion by roughly 5% to 7% in 2027.

Less visible is the supply chain behind the chips. Broadcom helped design the TPUs and coordinates their production, while Google has also signed a deal with Marvell Technology that includes some TPU work and MediaTek also does work on the program. Vahdat declined to comment on their roles. OpenAI and Anthropic both run in-house chip efforts staffed with former TPU engineers, and a startup founded by two ex-TPU engineers is developing a chip for long-context inference that is at least a year from finishing design.

Sources: Yahoo Finance, Bloomberg, fool.com.