AI Daily Brief: Google's New Flash Models Land as AMD and Anthropic Circle a Chip Deal

Two things happened at once this week that usually don't: model releases got cheaper and the compute deals behind them got bigger. That's not a contradiction, it's the same price war showing up on both ends of the stack.

Google quietly resets the price war

Google DeepMind shipped three new models overnight, Gemini 3.6 Flash, 3.5 Flash-Lite, and a security-focused 3.5 Flash Cyber, while conspicuously saying nothing new about the 3.5 Pro flagship developers have been waiting on. The workhorse Flash model now costs $1.50 per million input tokens, cuts token usage by up to 17% on typical tasks, and jumped from 37% to 49% on the DeepSWE coding benchmark, which is the kind of unglamorous efficiency gain that actually shows up in a startup's monthly API bill. Pair that with a Gemini 4 tease and a delayed Pro model, and the read is that Google is happy to let its cheaper models carry the news cycle while the frontier model stays in the oven longer than planned.

Anthropic locks up 3.5 gigawatts of TPUs

Anthropic also expanded its compute partnership with Google and Broadcom to 3.5 gigawatts of next-generation TPU capacity coming online from 2027, tripling what Broadcom was already supplying this year. The company said its annualized revenue run rate has passed $30 billion, up from roughly$9 billion at the end of 2025, which is the kind of growth that makes a multi-gigawatt hardware bet look less like overreach and more like catching up. Anthropic has been deliberately promiscuous about chips, running Claude across Google TPUs, AWS Trainium, and Nvidia GPUs rather than betting the company on one supplier, and this week's deal reads as the TPU leg of that hedge getting a lot heavier.

AMD opens its coming-out party

AMD's Advancing AI 2026 conference kicked off today in San Francisco, its biggest AI showcase since last year, with CEO Lisa Su's keynote landing tomorrow after a Monday that already saw Microsoft commit to deploying AMD's Helios rack-scale systems across Azure, joining Meta, OpenAI, and Oracle as early adopters. The event everyone's actually watching for is whether Anthropic shows up as a customer: analysts have flagged Anthropic hiring engineers with experience in AMD's ROCm software stack, and AMD reportedly listed Anthropic as a top-tier client internally alongside Meta, though nothing is confirmed until Su says it on stage. If it happens, Anthropic would be buying compute from Google, Amazon, Nvidia, and AMD in the same year, which tells you how uncomfortable every major lab has gotten with depending on a single chip supplier.

What it means for India

The compute arms race playing out between Google, Anthropic, and AMD this week is mostly a story about who gets squeezed once the chips actually ship, and Indian AI startups have been living that squeeze for months. Newer-generation Nvidia GPUs remain severely constrained even as older Hopper-era chips get easier to source, which is pushing up compute costs just as the IndiaAI Mission tries to hold the line with subsidized pricing, 34,000-plus GPUs offered to startups at roughly Rs 150 an hour. The one piece of good news actually points back to Google's release today: every time a Flash-tier model gets cheaper per token, it lowers the floor for what an Indian startup pays to build on top of it, which matters more to a bootstrapped Bengaluru team than another gigawatt deal in the US ever will.

Markets and AI money

The chip trade spent the day pricing in a deal that hasn't officially happened yet.

Name	Move / figure	Context
AMD	+4% this week	Rode the Microsoft Helios deal into its own conference
Broadcom	New multi-year backlog	Now supplying both Anthropic's TPUs and OpenAI's Jalapeño chip
Anthropic	$30B+ run-rate revenue	Up from ~$9B at end of 2025
Google Gemini Flash	$1.50 / million input tokens	Cheapest workhorse tier in the current price war

Broadcom is the name worth watching most: it's now the supplier behind Anthropic's TPU capacity and OpenAI's in-house Jalapeño chip, which makes it the one company getting paid regardless of which lab's compute strategy wins. Everyone else is placing a bet on an architecture. Broadcom's just selling the shovels to both sides of the race.

The pattern across today's news is the same one that's been building for weeks: model prices keep falling and compute commitments keep growing, and both trends are really the same bet that whoever locks in the cheapest, most diversified supply chain now gets to set the price everyone else competes against later.