Google's TPU 8tx: AMD CPUs inside, both new TPUs due 2027
Google added two systems to its eighth-generation TPU line on Monday, and neither one is a new chip. The interesting part is what Google put next to the silicon — AMD CPUs — and how long buyers will wait for it.
Google's eighth-generation TPU family grew by two systems on Monday: the TPU 8tx for confidential computing and the TPU 8i Max for memory-bound workloads. It is the company's fourth TPU system this year, and the first time it has introduced four in twelve months.
Neither is a new processor. Both are built on existing eighth-generation silicon with different packaging, networking and memory around it. Google describes the approach as deliberate, a way to deliver new hardware variants in months rather than years, and Amin Vahdat, the Google Cloud executive who announced the systems, told Bloomberg they are not redesigned chips but existing silicon reconfigured around specific customer requests. That framing matters: this is a portfolio-expansion announcement, not a performance leap. Google's April baseline for the 8t and 8i claimed up to 2.7x better training performance and 80% better inference price-performance than the previous generation, and nothing announced this week revises those numbers.
The 8tx is the more consequential of the two. It keeps data and model weights protected while they are being processed, not merely at rest or in transit, and it does that in hardware rather than through a software wrapper. Google says the design delivers up to twice the latency reduction of the TPU 8t with a near-zero performance penalty, and that a single system can fine-tune models, distill large ones into smaller ones and run reinforcement-learning training without moving sensitive data between clusters. That is the objection that has kept banks, hospitals and government buyers renting GPUs: confidential computing on general-purpose accelerators has historically cost throughput, so regulated workloads stayed on machines they controlled. Nvidia already offers a comparable confidential-computing mode on its GPUs, so Google is closing a gap, not opening one.
The detail worth circling is the host processor. Google is pairing the 8tx with custom AMD CPU hosts. The April eighth-generation systems ran on Google's own Axion Arm CPUs, which made a tidy story about vertical integration; this is the first AMD silicon inside a TPU system. Google has spent two years arguing that owning the whole stack is its advantage over Nvidia. Buying host CPUs from a third party puts a supplier inside the rack that Google does not control, and it is the kind of hedge a company makes when it expects demand to outrun its own fab allocations rather than the reverse.
The 8i Max attacks the other bottleneck. Agents working simultaneously across large codebases or hours of video stall on memory bandwidth rather than raw compute, so Google put two 8i chips on a single package. The result, per Google, is double the compute, double the SRAM and double the HBM bandwidth within the same rack footprint — significantly more performance per square metre of data centre floor. Note what the company does not claim to have doubled: HBM capacity. The two-chip package widens the road across the memory wall; it does not remove the wall.
Timing is the part Google left out of writing. Its announcement carries no availability date and no price for either system, and no eighth-generation TPU has a published price. Vahdat told Bloomberg both new versions are expected to launch next year, meaning 2027, with Oracle's cloud expected to start offering TPUs the same year. DataCenterDynamics independently reports the same specifications and the same silence on dates, which is the tell: the hardware is real, the schedule is a year out, and the announcement is aimed at buyers planning 2028 capacity now.
The commercial context explains why Google is announcing packaging variants at all. It says it now has nearly 100 external TPU customers, spanning frontier labs, trading firms, retailers and universities, and that every TPU it can supply is already booked. Bloomberg Intelligence estimates TPU sales will pass $135 billion by 2028, from roughly $12 billion this year excluding TPU revenue booked through Google Cloud, and puts TPUs about 30% cheaper than Nvidia GPUs at comparable throughput. Apple, Meta and Oracle are named customers; Jump Trading disclosed buying "thousands of TPUs" in Alphabet's third-quarter filing; Anthropic holds an option for as much as $150 billion of TPUs between 2028 and 2031, according to Bloomberg.
The counter-case is not complicated. A system that arrives in 2027 does not compete with what Nvidia ships in 2026, and Google's capacity is already spoken for, so external buyers are queueing behind Google's own models and Anthropic's option. Nvidia answered the shift toward inference by acqui-hiring the Groq team and licensing its inference hardware — a cheaper move than a new chip line, and one that lands this year. Google's advantage is price per token at scale for customers willing to move their PyTorch workloads onto a second stack; its disadvantage is that the second stack now arrives on a schedule Google will not print.
CoreWeave reports third-quarter results on Tuesday after the close with its debt trading like a distressed credit rather than an AI growth stock.
Yields on its 9.625% unsecured notes due 2032 pushed above 16% late last week, and the spread on five-year credit default swaps peaked near 1,289 basis points on November 5 before easing to roughly 1,155, CNBC reports. That is insurance against default priced for something close to a coin flip: about a 63% implied chance of default within five years, up from near 30% a year ago, according to Reuters. The stock is down roughly 49% in a month.
The balance sheet explains the repricing. CoreWeave carried $35.6 billion of borrowings at the end of June, $51.6 billion including leases, against $5.5 billion of cash. Its own guidance points to third-quarter interest expense of $860 million to $940 million against adjusted operating income of $200 million to $260 million — an interest bill running three to four times operating profit before a dollar of capital spending. Consensus expects revenue of about $3.54 billion, up 160% year on year, with Microsoft still supplying the bulk of it: about 67% of 2025 revenue and 36% in the second quarter.
We covered the mechanics when CoreWeave raised $3 billion more and its debt passed $35 billion. The difference now is directional: the market has stopped questioning the backlog and started pricing the borrower. Freedom Capital's Paul Meeks put it plainly to CNBC — a "doom cycle" in which creditors worried about repayment demand higher rates, which makes the next borrow more expensive, which makes the worry worse. IREN offered the counterpoint on Friday: it said 48% of its backlog is tied to Microsoft, then jumped 20% after AI cloud revenue more than doubled.
What to watch: whether CoreWeave's management can describe a financing path that does not depend on a beaten-down share price, Nebius's numbers on Thursday, and whether the 8tx's confidential-computing pitch lands with the regulated buyers who have so far stayed on GPUs.
If your company buys AI compute, have the credit markets changed how you plan capacity for 2027? Write to us via our contact page.
