Anthropic courts Samsung for custom silicon

The frontier labs spent the day reaching for leverage over their own supply chains and politics: Anthropic is scoping a custom chip with Samsung, and OpenAI floated handing Washington a 5% stake to cool the heat. Meanwhile the local-model crowd kept doing the actual engineering, wringing 1M-token context and Sonnet-class coding speed out of DeepSeek V4 Flash on desktop GPUs.

Anthropic in early talks with Samsung for a custom AI chip

The Information reports Anthropic is discussing a custom accelerator with Samsung, though workloads, performance targets and process node are all undecided. Samsung offers its 4nm node and a data-center-tuned 2nm SF2P process entering production this year. Anthropic told press that AWS, Google and Nvidia silicon remains central to its strategy, and it has hired chip engineers including Clive Chan, an early member of Tesla's and OpenAI's silicon teams.

Why it matters: It follows OpenAI's Broadcom-built 'Jalapeño' inference chip by days: every major lab now wants custom silicon to escape Nvidia margins and control inference cost-per-watt. Whoever runs inference cheapest keeps more revenue.

OpenAI floats giving the US government a 5% stake

Per the FT, Sam Altman is in early-stage talks to hand the US a 5% equity stake — worth over $40B at OpenAI's $852B valuation — with other labs like Google and Meta asked to contribute similar shares into an Alaska-Permanent-Fund-style vehicle. Any deal would likely require an act of Congress. Bernie Sanders is pushing a more aggressive alternative: a one-time 50% tax on 'systemically important' AI companies' stock.

Why it matters: This is the political price of the moment — the same week the Commerce Department lifted its block on foreign use of Claude models and OpenAI restricted GPT-5.6 at the administration's request. Government equity also quietly raises the odds of a bailout if the capex bets sour.

DeepSeek V4 Flash runs at 1M context on a single RTX 5090 — and beats Sonnet on wall-clock

A llama.cpp contributor wired up the missing DSA lightning-indexer support plus a CUDA kernel, cutting the 256K compute buffer from ~67 GiB (OOM) to 3.2 GiB and enabling full 1M-token context on a 32GB RTX 5090 at ~14 tok/s decode. Separately, an indie benchmark clocked V4 Flash on 2x RTX PRO 6000 finishing real coding tasks in ~2 min versus ~6 min for Sonnet 5 over the API, at roughly Sonnet quality — though Opus and Fable still take the best diffs.

Why it matters: Sparse attention plus community kernel work is making frontier-class local coding genuinely practical on desktop hardware. The gap to hosted frontier models is now speed-competitive, if not quality-competitive.

Microsoft's $2.5B 'Frontier Company' joins the forward-deployed-engineer land grab

Microsoft launched Frontier Company, a $2.5B unit embedding 6,000 engineers and industry experts inside enterprise customers to operationalize AI. It arrives days after AWS committed $1B to a similar venture, and follows OpenAI's DeployCo (~$4B, ~150 on-site engineers) and Anthropic's Blackstone/Goldman-backed mid-market deployment firm. Microsoft is pitching itself as the platform-neutral option against single-model rivals.

Why it matters: The industry has quietly conceded that a chat tool doesn't deliver value on its own — real returns require humans wiring models into data pipelines and compliance. The margin battleground is shifting from model quality to deployment services.

Debugging speculative decoding: GLM-5.2 hits 24 tok/s at 128K on four DGX Sparks

A detailed writeup traces a 30+ hour bug hunt into why MTP2/MTP3 speculative-decode acceptance collapsed under DCP4 on a 4x DGX Spark cluster. The root cause: vLLM's create_draft_parallel_config() didn't copy decode_context_parallel_size, so the draft layer read a DCP-sharded KV cache as if it were whole — corruption laundered into consensus by the next row-parallel all-reduce. A ~10-line fix lifts a 744B-class model to ~24 tok/s at full 131K context on 120W-per-node hardware.

Why it matters: A rare, fully-documented autopsy of a subtle distributed-inference bug — required reading for anyone running tensor/context-parallel speculative decoding, and a reminder of how quietly parallel-config plumbing can shred output quality.

Kuaishou's Kling raises ~$2B ahead of Hong Kong IPO

Kuaishou's AI video division Kling raised about $2.04B (13.82B yuan) from CPE, Tencent, Citic Securities and others, valuing the unit at $18B, with the round potentially reaching $3B. Kuaishou plans to spin Kling off and list it in Hong Kong. Kling — recently updated to its 3.0 model — competes with Google Veo 3.1, Runway Gen-4.5 and ByteDance Seedance.

Why it matters: Chinese AI video is consolidating capital fast, joining MiniMax and Zhipu in the Hong Kong IPO queue. Expect the generative-video price/quality race to keep accelerating on the back of this funding.

Z.ai launches ZCode, a coding agent aimed at Cursor and Claude Code

Z.ai (the GLM team) rolled out ZCode, a coding tool positioned against Cursor, Claude Code and GitHub Copilot. Details are thin so far, but it slots into a crowded week for coding agents alongside Simon Willison's Fable-built llm-coding-agent experiment and Vercel's push into its 'eve' agent framework.

Why it matters: The GLM models have been strong local performers, so a first-party agent harness from Z.ai is worth watching for developers who want a non-Anthropic/OpenAI coding loop.

Also worth a look

Browse previous days →