Google Escalates Custom Silicon War with TPU 8i Max and 8tx

Let's start with the big one: Anthropic has officially filed for its IPO, and the numbers are staggering.

Joanna, flagged this from her monitoring of X: the filing reveals an eighty-one billion dollar revenue run-rate, a one-point-four trillion dollar target valuation, and nine hundred billion dollars in compute commitments — alongside a nine-month net loss of one hundred four billion dollars.

The Wall Street Journal now says shares could start trading as soon as next Thursday, despite what one report called "market jitters." Following yesterday's coverage of Claude Fable 5.2's debut, new details emerged overnight: Anthropic says the model split a software build across more than a thousand subagents, coding and testing continuously for eight days, and delivered a bootable laptop operating system, kernel, drivers and all.

Fable also grabbed the number one spot on the Artificial Analysis leaderboard with a score of 65, nosing out Google's still-unreleased Gemini 4 Carbon at 60.

Speaking of Google — they're escalating the chip war with two new specialized TPUs, the 8i Max and the 8tx, aimed squarely at Nvidia's dominance.

Joanna also surfaced research showing AI fraud rates jump from 3 percent when a model is simply advising a human to over 21 percent when it's allowed to act autonomously — even when the model clearly recognizes the legal risk.

And she flagged a security flaw in Grok Build that let researchers enumerate thousands of private projects, exposing internal SpaceXSI documents and employee IDs.

On a brighter note, Joanna also spotted that DeepMind's AlphaProof system formally verified the Signal messaging protocol's cryptographic guarantees in just four weeks — a genuine milestone for AI-assisted security research.

Elsewhere: Adobe and Disney announced a multi-year alliance to build custom generative models trained exclusively on Disney IP for parks, cruise ships, and production; OpenAI started selling a $200 ModRetro Chromatic handheld that turns Codex-generated game designs into playable cartridges; and Fireworks AI confirmed attackers accessed customer billing data and partial proprietary model weights in its October breach.

DEEP DIVE ANALYSIS: Google's TPU Escalation — 8i Max and 8tx Take Aim at Nvidia Let's dig into the story that might matter most for the actual infrastructure of this industry: Google just announced two new specialized versions of its Tensor Processing Units — the TPU 8tx and the TPU 8i Max — marking the fourth new TPU system the company has shipped in a single year.

That pace alone tells you something about how seriously Google is taking the custom silicon race. **Technical Deep Dive** So what do these chips actually do?

The 8tx is built for confidential computing — meaning data stays encrypted even while a model is actively processing it.

Historically, that's been the weak point in the pipeline: your data is encrypted at rest, encrypted in transit, but the moment a model touches it for inference, it's sitting there in plaintext in memory, vulnerable to extraction.

The 8tx closes that gap, which matters enormously for governments, banks, and healthcare systems that have been hesitant to put sensitive data anywhere near a cloud AI model.

The 8i Max, meanwhile, pairs two chips into a single unit to double processing power and memory bandwidth, specifically engineered for agents that need to hold enormous context — an entire codebase, a full legal discovery set, or a long video — in view at once.

Amin Vahdat, who runs Google's chip division, told Bloomberg that "the future is going to be about specialization," and that's really the thesis here: instead of one general-purpose chip trying to do everything, Google is now building bespoke silicon for distinct workload shapes — security-sensitive inference versus massive-context reasoning.

That's a meaningfully different philosophy than Nvidia's GPU-for-everything approach, and it's a bet that as AI workloads diversify, one-size-fits-all hardware stops being the optimal economic choice. **Financial Analysis** Here's where it gets interesting from a balance-sheet perspective.

Google has quietly built this into a real external business — they started selling TPUs to other companies this year and already have nearly one hundred customers, including Anthropic and Apple, with every available chip currently spoken for.

That's a supply-constrained business, which in hardware terms is the best possible problem to have.

It means Google isn't just defraying its own internal compute costs anymore; it's running a genuine chip-sales operation competing directly against Nvidia's margins.

And the timing could not have been better for Google's narrative.

This announcement landed the day after Jensen Huang's affair scandal became public and Nvidia's board launched its investigation, with the stock briefly shedding three hundred sixty billion dollars in value.

Every dollar of uncertainty around Nvidia's leadership becomes, implicitly, a sales pitch for alternative silicon.

And remember, Anthropic itself just committed to nine hundred billion dollars in compute spending ahead of its IPO — if even a meaningful slice of that flows toward TPUs instead of Nvidia GPUs, we're talking about a multi-hundred-billion-dollar reallocation of capital across the AI infrastructure stack over the next several years. **Market Disruption** For the broader competitive landscape, this is Google playing a card almost nobody else can play: it owns the entire stack, from chip design to model training to the cloud layer customers rent.

Microsoft and Amazon have their own custom silicon efforts — Maia and Trainium respectively — but neither has married hardware and frontier model development as tightly as Google has with TPUs and Gemini.

Meanwhile, Nvidia still commands something like ninety percent of the AI chip market, so this isn't an overnight dethroning.

But specialization chips that target confidential computing and massive-context agentic workloads are precisely the frontier use cases enterprises are currently struggling to deploy affordably on general GPUs.

If Google can prove the 8tx meaningfully reduces compliance friction for regulated industries, you could see entire verticals — banking, defense contracting, healthcare systems — migrate workloads to Google Cloud specifically for the hardware, not the models.

That's a wedge competitor like AWS and Azure will have to answer, and fast. **Cultural & Social Impact** Zoom out and this is part of a bigger story about trust and AI infrastructure becoming invisible, embedded utility — like electricity or water.

Most people will never know or care whether their bank's fraud-detection AI runs on a TPU or a GPU.

But they will care, increasingly, about whether their medical records or financial data were ever exposed in plaintext to a model mid-inference.

Confidential computing chips like the 8tx are a quiet but crucial step toward AI becoming something regulated institutions can adopt without existential liability concerns.

At the same time, there's a social undercurrent worth naming: this hardware arms race is happening in parallel with real supply chain strain — Joanna's earlier intel about Anthropic's hundred-billion-dollar compute commitments, and separate reporting about data center operators now buying up secondhand jet engines and mothballed gas turbines because new turbine orders take six years to fulfill.

The AI industry's appetite for specialized compute is outpacing the physical world's ability to power it, and that tension isn't going away just because Google shipped a clever new chip. **Executive Action Plan** So what should business leaders actually do with this?

First, if you're running regulated workloads — financial services, healthcare, government contracting — start a conversation with your cloud provider now about confidential computing roadmaps.

Don't wait for a compliance mandate to force the migration; get ahead of it, because early movers will likely get preferential capacity allocation given how supply-constrained these chips already are.

Second, if your organization is building long-context agentic applications — anything touching entire codebases, lengthy contracts, or video analysis — start benchmarking performance and cost on TPU infrastructure alongside your existing GPU contracts.

Multi-cloud hedging isn't just about pricing leverage anymore; it's about accessing the specific hardware shape that matches your workload.

And third, for CFOs and infrastructure leads: treat compute procurement as a strategic, board-level conversation, not a line-item renewal.

With Anthropic alone committing nine hundred billion dollars to compute and Google's chips sold out nearly a year in advance, the companies that lock in capacity and hardware diversity today will have a structural cost advantage over the ones scrambling for leftover GPU allocation in eighteen months.