Overview
The day’s AI talk sat between giant infrastructure cheques and practical agents that can now do recognisable office tasks. NVIDIA was tied to both the capital race and open security work, while founders argued over open weights, indie software, and the hard limits of running giant models. Away from AI, Alphabet’s SpaceX stake, Uber’s Kalanick bet, and the Basecamp founders’ long game added some useful texture.
The big picture
NVIDIA may underwrite OpenAI’s next compute bill
Bloomberg reports that NVIDIA is in talks to provide a guarantee of about $250bn to help OpenAI lease compute from SoftBank’s planned $500bn data centre project in Ohio. The site is expected around 2028 and could draw 10 gigawatts of power, which gives a sense of the scale now attached to frontier AI.
This would put NVIDIA in a role beyond chip supplier. It would be helping finance the demand that keeps its GPUs central to the AI build-out, though the discussions are still early and could change.
Agents start to look like office colleagues
Sam Altman showed ChatGPT Work taking a single phone prompt and turning it into a full group trip planning flow: checking past chat history, suggesting options, building a coordination site, handling decisions and reservations, then drafting the email. Tibo pushed the same idea into daily admin, from bill negotiation to clearing spam subscriptions.
Riley Brown’s Claude Excel demo sat in the same category. A single prompt led to a multi-tab NVIDIA financial model with public sources, live formulas, checks, forecasts, valuation work and charts. The common thread is not chat, it is finished output.
Open weights become the line in the sand
Guillermo Rauch said Vercel has signed the Open Weights and American AI Leadership letter, joining a growing group arguing that open model weights are the next step for open technology. NVIDIA also launched the Open Secure AI Alliance with Microsoft, IBM, Cisco, Hugging Face, SpaceX and others, aimed at shared tools and research for safer AI software and agents.
The politics around this are getting sharper. Anthropic is being cast by critics as the outlier, with its safety-first stance clashing with a broader industry push for open access, transparency and American competitiveness.
Kimi-K3 wins the arena, but the hardware bill bites
Kimi-K3’s rise to the top of the Frontend Code Arena was impressive, but @0xSero focused on the practical pain. At 2.8 trillion parameters, the raw BF16 footprint would be 5.6TB, with even aggressive tricks still leaving it in the realm of hundreds of gigabytes of VRAM and system memory.
The complaint is plain: benchmark wins matter, but models that are open in theory can still be out of reach for builders without serious hardware.
Indie software has a new traffic problem, not a death notice
Greg Isenberg pushed back on the idea that AI is eating indie software. His point was that a certain playbook is under pressure: make thin SEO pages, rank on Google, and collect free signups. If ChatGPT now answers the question directly, that channel is weaker.
But he points to Stripe Atlas data showing more new companies, faster early revenue, and more solo founders. The builders doing best are not just cloning old SEO tactics, they are using AI to make smaller teams move faster.
Alphabet’s old SpaceX bet now looks enormous
Whole Mars Catalog highlighted Alphabet’s disclosed $94.1bn stake in SpaceX after the company’s IPO. The stake dates back to Google’s 2015 investment, when SpaceX was valued at around $12bn.
With SpaceX now marked near a $1.5tn market cap, that old bet has turned into a major asset. Much of the holding remains under post-IPO sale restrictions, so it is not instantly liquid, but the paper gain is hard to ignore.
Uber backs Travis Kalanick’s robotics comeback
TechCrunch reported that Uber is investing $100m in Atoms, the robotics company led by former Uber CEO Travis Kalanick, as part of a $1.7bn funding round. The move reunites Uber with Kalanick eight years after his exit from the company.
Atoms is aiming at industrial AI and physical automation, including mining and transport autonomy. It has also acquired Anthony Levandowski’s Pronto, adding another familiar name from the self-driving car era.
Long-running agents need a proper memory trail
Yohei Nakajima argued that anyone building long-running agents will end up using immutable event logs. The reason is simple: when an agent runs for hours or days, you need a clear record of what happened, not just a snapshot of the final state.
He compared it to git histories, accounting ledgers, medical audit trails and flight recorders. For agents, that kind of record makes debugging, recovery and trust far easier.
Science fiction missed the plain weirdness of today’s AI
Ethan Mollick noted that science fiction did not really anticipate the AI we have now. It imagined conscious machines, vast superintelligences and robots with inner lives, but not statistical language models that can write code, tutor students, draft strategy and run small workflows.
That mismatch is part of what makes the present feel strange. The tools arrived through data, scale and prediction, not through the classic story of a machine waking up.
Basecamp’s founders keep making the case for independence
DHH thanked Jeff Bezos for the confidence that came with his 2006 minority investment in the company behind Basecamp. The deal gave DHH and Jason Fried more security without forcing them into the usual venture path of chasing rapid growth or an exit.
Jason Fried also praised David Senra’s long interview with DHH, which circled many of the same themes: constraints, remote work, profit, saying no, and keeping control over product decisions.
