Doomcoding Doomcoding is replacing doomscrolling by directing your agents to code from your phone and scrolling between agents working on multiple PRs. Antidote to doomscrolling as dopamine hits come from reviewing output of work. @karpathy - Good addition to AI vocabulary?

I spent the week at the AI Infra Summit in Santa Clara. Pretty uneventful conference overall, but the narrative continues to accelerate. Kind of beating a dead horse here but here you go… Eight more neoclouds every hour on the hour. These are financing vehicles housing the debt for the l labs and the hypers, but it’s still just open AI and Anthropic under the hood driving much of the consumption. Offtake owns the market. $NVDA orchestrating this giving one year backstop + 20% operating cushion and now buying up colo capacity so the financing vehicles can build the neo cloud via colo. really masterful GTM. predatory but masterful. The pace of this is growing faster and faster. NVDA hangs this out there with the requirement that has to be deployed by X date (so they get revenue recognition fyi) and if you can’t hit that date, you lose the allocation. This is causing quite the acceleration. The big Jensen clouds are only working on labs it seems as they got first dibs and are not taking any other business. Aka Lambda. Notice they killed their token factories for the most part. It will be interesting how that plays out although the founders will have already taken billions off the table. I don’t think anybody cares about Neo sustainability, except for maybe $NBIS. That lack of a concern for LT sustainability is a bright red flag. A few other neos doing some cool things though so I don’t wanna be totally dismissive there. Lots of bare metal only though because that’s the lab desire $AMD using similar tactics with less scale and interest because Nvidia rules the market. AMD seems more choked on supply chain. They’ve downgraded a couple clouds I know from current to older gen. Fireworks, baseten, together remain the other critically important demand side players. If anyone thinks these are some lesser clouds because they’re not public guess again. this is the layer two power in the market with Nebius. This group is thankfully seeing more logos. that is welcome news because, guys, this market right now a jenga tower with two pieces keeping the entire industry afloat at the bottom. do not lose sight of that Jensen playing a game to subsidize the labs until the compute catches up and gross margin can be positive including training because it’s really dumb to exclude that. It’s clearly bloat on training thats causing the FCF concerns. Inference absolutely prints money today on blackwells. Compute buyers are 99% neo clouds right now for non $NVDA/ $AMD . Can’t break into the hypers because they don’t want to manage an ASIC and all they care about is cost per mill. enterprises are still mostly via cloud. neos driving the market forward will buy any and all compute they can find and pre-sell. They’re getting about 30% down on the pre-sale for a deployment six months out. They transact this at whatever the minimum debt coverage ratio is on their loan facility is to build as much scale as they can quickly. They will pay $NVDA full list with a smile on their face. The old timer DC guys think it looks like the Internet roll out with the financing vehicles, but those with the more optimistic / those that understand the impact of coding and how it drives compute consumption have a different view and remain fantastically bullish. I am one of those people. if we get one more use case, it’s all to the moon. Anecdotally if OpenAI really submitted a Nobel level response to the math problem it’s a game changer for everything. Frontier goes to something else. Open source explodes for cheaper coding because why would you let your India back office develop some stupid feature using a Nobel winning model and incur all that cost? Muse is really goos and while Claude and openai anthropic could’ve nailed it in a second, they’re too focused on AGI to do something so simple and practical, and I think it’s great for the market that meta really nailed something. Muse gonna blow up fast.

Not enough people are talking about CXMT's move into NAND. China's DRAM champion is setting up an R&D pilot line for NAND in Beijing. Still early stage and a few year away from mass production. But the long-term structural implication for the memory market is significant. Some history for context. Elpida went bankrupt in 2012, and that event consolidated DRAM down to three real players: Samsung, SK Hynix, and Micron. That scarcity was a big part of why DRAM profitability improved so much over the following decade. NAND took a different path to consolidation. Toshiba's flash business spun out as Kioxia, Western Digital merged with SanDisk, and SK Hynix bought Intel's NAND unit to form Solidigm. Same outcome: fewer players, better pricing discipline. Now the events are happening in reverse order. YMTC already is a solid NAND player. CXMT makes the market crowded, it reopens a market structure question the industry thought was settled.

Ridiculous fear mongering. Why increase CPU temperature, just blink your LED twice when there is a problem. On a serious note, when sensational story telling in a domain gets to this level, next step is usually denial followed by disappointment.

What doesnt kill us makes us stronger. Plugin4Shell hits major coding agents. SHA-pinning on Claude Code, Codex, Gemini CLI and Copilot: a hijacked or swapped plugin can auto-update into remote code execution. Anthropic and OpenAI patched; Google will not patch deprecated Gemini CLI (migrate to Antigravity); Microsoft/GitHub dispute impact. Researchers used Claude to reach OpenAI’s private code: Hacktron used a cyber-special Claude Opus build to exploit an OpenAI Discourse image bug, stole auth tokens that also worked on employee ChatGPT and GitHub, and opened a harmless PoC PR. OpenAI paid a $6,500 bounty.

Semiconductor equipment subsystem lead times: Deposition-equipment component lead times ~4 → up to 10 months. Laser-process Japanese parts up to 40 months, no premium pulls them forward. Packaging-tool components lead times up +~50%. Fab tool deliveries ~2×. Drag already flagged on Samsung Pyeongtaek and SK Yongin.

Nexperia announced a partnership with Tata Electronics. The Dutch power/discrete chipmaker will make MOSFETs and power-control chips at Tata’s ~$11B 300mm fab in Dholera, Gujarat, and do assembly/test at Jagiroad, Assam. Why it matters: • Tata’s delayed fab gets a named customer path • Nexperia is owned by China’s Wingtech, yet a Dutch court stripped Wingtech of control after a 2025 state intervention meant to stop ops moving to China • Beijing had hit back with a Dongguan packaging export ban that starved carmakers of chips • Thursday’s Tata deal was endorsed by Nexperia’s current Dutch management, further sealing the split

Nvidia’s next AI rack just posted its first official MLPerf scores. Vera Rubin NVL72 (preview) vs today’s GB300 rack: • up to 3.7× on Qwen3-VL • up to 2.5× on DeepSeek-R1

If SHA256 breaks tomorrow, the internet stops working. That's not an argument against quantum computing. It's an argument for better cryptography. When cloning was first possible, people warned of evil armies of clones. Thirty years later, cloning gave us better cancer research and zero clone armies. Y2K had the world convinced planes would fall out of the sky at midnight. The fix wasn't "stop building software." It was "go fix the code." You don't slow down technology because it can be misused. You use the technology to build stronger infrastructure around it. Technology is a neutral force. It cuts both ways. I'm betting on the side that builds, not the side that's scared of its own shadow. "Pace the frontier" sounds like wisdom. History says it's usually just fear.