Amazon in talks to buy Decart for about $7 billion
A quiet-looking Monday has one big rumor anchoring it: Amazon may be about to spend roughly $7 billion on the chip-porting startup Decart — after Anthropic walked away from the same deal. Meanwhile Intel spelled out its inference-GPU answer to Nvidia, and OpenAI quietly handed Codex and ChatGPT Work users a much roomier default.
Amazon is reportedly in talks to buy Decart for about $7 billion. The Wall Street Journal, citing people familiar with the matter, reported Sunday that Amazon is negotiating to acquire the San Francisco AI startup — a deal that could be agreed soon barring a collapse in talks or a rival bidder, and that would rank among Amazon's largest acquisitions ever. Neither Amazon nor Decart has commented, so this stays a rumor until paperwork says otherwise. The backstory matters more than the price tag: Decart builds the software layer that lets developers run AI models across different chip vendors, plus the real-time world models Lucy and Oasis, and Anthropic reportedly walked away from a roughly $6 billion deal for it last month after due diligence. For Amazon the logic is Trainium — AWS already gave Decart early access to its Trainium3 chip, and Decart's porting work is exactly the kind of software that makes non-Nvidia silicon usable for customers who don't want to be locked in. A ~$7 billion valuation on ~$400 million raised and a May round near $4 billion is startup math at full stretch, but if it closes, it's a clear signal that the fight for AI compute is moving from who has the chips to who can make any chip work.
Intel details the D270, a 480GB air-cooled inference GPU. Intel's new data-center GPU packs up to 480 GB of LPDDR5X memory per card, 32 Xe3P cores and 256 matrix engines into a 350W air-cooled PCIe card — with roughly 1.3 PFLOPs of FP8 compute, per Intel's announcement as covered by WCCFTech. On paper the memory capacity tops both AMD's MI450X (432 GB) and Nvidia's Vera Rubin (288 GB), and Intel's pitch is honest about where inference actually hurts: long contexts and agent sessions keep models resident, so capacity, not just bandwidth, decides how many users a rack can serve. The caveats are real, though — there's no price, no on-sale date, independent benchmarks don't exist yet, and the "beats RTX Pro 6000" claims are Intel's own internal analysis from February. Customer sampling is targeted for the second half of 2026, and the clock on that window is ticking.
OpenAI lifts GPT-6.1 Sol's default context to 500K tokens. OpenAI head of product Tibo Sottiaux announced the change as day seven of its shipping sprint: the default context in Codex and ChatGPT Work rose from 272,000 to 500,000 tokens, and the long-context surcharge under ChatGPT subscription quotas is gone. The honest fine print: the model's maximum context was already 1,050,000 tokens at launch, and the API is untouched — prompts above 272K still pay the 2x input, 1.5x output multiplier. So the headline story isn't a bigger ceiling, it's OpenAI moving to actually fill one on subscriptions — a sign that long-context capacity is now a retention lever, not a pricing tier.
What to watch: whether any rival bidder surfaces for Decart before Amazon signs — Nvidia reportedly held advanced talks earlier.
Is the chip-porting layer the real AI arms race, or is this Amazon overpaying for insurance against Nvidia? Tell us in the comments.
