THE 60-SECOND VERSION
● NVIDIA Groq 3 LPX in full production, up to 256 per rack, claiming 4x faster agent responsiveness.
● OpenAI benchmarked Jalapeno against GB300 the day after LPX shipped, and did not test against Vera Rubin.
● The deeper story: chips are being redesigned around agent loops rather than batch throughput.
● 4 days to the 31 August double deadline.
NVIDIA shipped a dedicated inference chip
Groq 3 LPX entered full production on 24 August at Hot Chips. It extends Vera Rubin NVL72 with up to 256 accelerators per rack, splitting the workload — Rubin GPUs handle large-scale context processing while LPX handles latency-critical decode. NVIDIA claims 4x faster responsiveness for agents than the nearest alternative platform, and demonstrated 3,400 tokens per second on Gemma 4 31B at 100K context. Nebius is the first AI cloud to adopt it.
The timing next to OpenAI is the part worth noticing. Jalapeno's benchmarks, published the following day, compared against GB200 and GB300 — the previous generation — and did not test against a Vera Rubin configuration with LPX doing exactly the decode work Jalapeno targets.
Full breakdown: what LPX does, and the claim to be careful with →
And the design shift underneath both
NVIDIA also detailed the Vera CPU: 88 custom Olympus cores across six chiplets, deliberately prioritising single-thread performance for orchestration and tool-calling over raw core count. That is the opposite of a decade of server CPU design, and it follows from what an agent loop actually is — a serial chain where each step waits on the last.
Three designs, one assumption. LPX splits decode from context. Vera optimises orchestration. Jalapeno keeps model state local to cut data movement. All three assume the workload is an agent iterating rather than a batch job completing.
Why this follows from the Linear and Salesforce agent data →
Claude got a third browser
Anthropic shipped a Chromium browser inside Claude Cowork on 26 August, rolling out to all paid plans over the following week. No extension, nothing to install — it opens in the side panel when a task needs a website, and Claude reads, clicks and types in it.
The property that matters is that it runs a clean profile with none of your logins. That makes it the safer option for anything not requiring authentication, because a malicious page cannot reach sessions that are not loaded. Anthropic says the Chrome extension remains the right choice when the work is on a page you already have open and signed into.
Read alongside Cloudflare shipping Kitesurf three weeks ago, the convergence is the story: two companies independently decided agents should not use your browser, and reached opposite engineering answers. Cloudflare deleted Chromium; Anthropic embedded it.
All three Claude browsers compared, and which to use when →
Also this week
- Thomson Reuters launched its own frontier model for legal work.
- Mac mini demand has spiked partly on buyers running models locally, with lead times stretching weeks to months. Apple plans to move production to its Houston facility.
- CoreWeave deployed Spectrum-X Multiplane into production, connecting Vera Rubin racks across parallel switches.
Four days to the double deadline
| Date |
What happens |
| Aug 31 |
Claude Sonnet 5 goes $2 to $3 per million input, $10 to $15 output, plus a tokenizer change adding 10 to 35 percent tokens on code |
| Aug 31 |
GPT-5.4 and GPT-5.4 mini leave Codex for ChatGPT sign-in users. Still available via API key |
| Aug 31 |
kimi-k2.5 and moonshot-v1 sunset, migrate to kimi-k3 |
| Early-mid Sept |
Grok 4.7 window per Musk. Still no model ID at docs.x.ai |
| Oct 1 |
OpenAI vs Apple hearing |
| Oct 24 |
deepseek-chat and deepseek-reasoner deprecated |
| ~Nov 21 |
GPT-5.6 Sol promotional pricing ends, $4/$20 back to $5/$30 |
Still open
- Reddit's ChatGPT citations. Down 86 percent in three days after an unannounced retrieval change, with the share moving to vendor documentation. What publishers should check →
- CVE-2025-62593 in Ray. Anything below 2.52.0 remains exposed. What to check →
- OpenAI's public S-1. Filed confidentially 8 June, still not on EDGAR.
FAQ
What is NVIDIA Groq 3 LPX?
A dedicated interactive inference accelerator extending Vera Rubin NVL72, in full production since 24 August 2026, with up to 256 accelerators per rack. Rubin GPUs handle context processing; LPX handles latency-critical decode.
Did OpenAI beat NVIDIA?
OpenAI claims Jalapeno delivers 1.5 to 1.9 times more work per watt than GB200 and GB300 — the previous generation. It did not test against Vera Rubin, and both sets of figures come from the companies selling the hardware.
Why are chips being redesigned for agents?
An agent loop is serial: decide, call a tool, wait, read, decide again. Adding parallel cores does not speed up dependent steps, so latency per step matters more than aggregate throughput.
What changes on 31 August?
Claude Sonnet 5 rises to $3 in and $15 out per million with a tokenizer change alongside, and GPT-5.4 and GPT-5.4 mini leave Codex for ChatGPT sign-in users. Both remain available via API.