Cornelis raised $205 million. Its pitch: the network, not the chip
Monday's news ran from the inside of the AI rack to the inside of the White House — a networking challenger that says the GPUs are not the bottleneck, a private Altman-Trump meeting that surfaced the same day the two split publicly on slowing down, and a mathematician's attempt to say what the profession should be for.
Cornelis Networks announced roughly $205 million in funding alongside Active Compute Fabric, an open networking architecture that puts programmable compute inside the fabric itself. The company, which spun out of Intel in 2020 and is led by CEO Lisa Spelman, says the round led by IAG Capital Partners will pay for a scale-up networking buildout, its next-generation scale-up and scale-out roadmap, and manufacturing and customer deployments of its CN6000 hardware. The product is aimed at a specific waste: as models and clusters grow, communication overhead and synchronization leave expensive accelerators sitting idle, and compute, storage and memory have all become workload-aware while networking still mostly just moves bytes between endpoints. Active Compute Fabric is meant to change that — lossless transport plus in-fabric acceleration that adapts to the workload, offloads collective operations and works on data while it is in flight, built on the company's third-generation switch silicon and preserving accelerator choice through open industry standards. Cornelis also said Qualcomm Technologies will collaborate on validation toward future rack-scale AI data-center designs, with Qualcomm's data-center chief Tony Pialis joining Spelman at AI Infra Summit. The strategic read: Nvidia's chips technically run on other fabrics but are tuned for Nvidia's own software stack, so the fastest route to diluting that lock-in is not a rival GPU — it is making the plumbing neutral. Cornelis says the fabric is already shipping, with a new generation later this year. This sits in the same vein as our earlier look at interconnect money — Ayar Labs adds $150M more as copper interconnects hit the wall.
Sam Altman asked for a private meeting with President Trump backstage at the Republican midterm convention last Thursday, according to three people familiar with it, and the account surfaced Monday alongside a sharp public split between the two men. Altman requested the sit-down, and only a brief overview of the discussion — centred on artificial intelligence and its growing power — was shared with MS NOW, which reported the meeting; Reuters reported the same account and said it could not immediately verify it, and neither the White House nor OpenAI responded to requests for comment. The timing is the story. Days earlier, Altman posted that slowing the pace "will be well worth this cost; no amount of American competitive pressure should justify recklessness, or let capabilities get ahead of alignment and monitoring," while Trump spent Monday calling guardrail demands a hoax and saying the only control AI needs is a strong president. The reporting also fills in a gap this site has tracked: the voluntary framework created by a June executive order, which gives the government access to new models for up to 30 days before release, was completed on its August 1 deadline and discussed with companies the next day — Anthropic attended, no readout was released, and the framework has still not been made public. We covered the president's answer to the pacing ask earlier today — Trump's answer to the pacing ask: he already has criminal power over the labs.
A University of Toronto mathematician published a concrete reform plan for his profession on Monday, and its sharpest proposal is that the PhD should be graded on the defence, not the paper. Writing in an essay titled "A beginning for mathematics," Daniel Litt takes as his premise that systems robustly superhuman at mathematics are close — three years ago they could not reliably add two numbers, a year ago internal models at OpenAI and DeepMind reached IMO gold-medal standard, and now they resolve major open questions — and argues the institutions cannot survive unchanged when "anyone with a laptop and a few hundred dollars can generate what would have been an Annals paper last year." His fix is to stop using mathematical text as a proxy for mathematical understanding: accept results regardless of provenance, but award degrees on a rigorous oral defence of a deep topic, privilege talks and sustained discussion over papers, interview graduate applicants the way departments already interview faculty hires, and reward people who build research programmes and seminar cultures rather than theorem counts. He is explicit that none of this depends on the labs behaving, or on capabilities disappearing: "AI does not care if you are anti-AI." It is the most institution-facing answer yet to a problem we examined structurally — Math's credit system was built for humans. AI just broke it.
What to watch: whether the White House confirms the convention meeting at all, and whether Cornelis prices the fabric as a per-rack product or a subscription to an alternative stack.
If a PhD were awarded purely on a live defence, would that raise the bar for human understanding or just hand the advantage to whoever performs best under pressure? Tell us in the comments.
