Wink Pings

Model Companies Are All Building Chips—Don’t Rush to Write Nvidia’s Eulogy

A tweet from Harry Stebbings laid out the math behind model vendors developing their own chips. The cold water in the replies is more worth reading: TSMC is still there, and the cycle won't move that fast.

Harry Stebbings tweeted a question: why is every model provider moving toward custom silicon? In the video, @goodwin_ml lays out the math clearly: when serving frontier models, margins will be squeezed by the high gross margins of incumbent chip vendors; first-party chips are a way to preserve a buffer. Original tweet

The replies under this tweet are more interesting than the main tweet. One reply reminds people that TSMC is still making chips for most vendors, and that profit won't disappear into thin air. Custom chips can change design, the software stack, and bargaining power; manufacturing is still another business.

Apple is often used as an analogy. In the replies, someone says it took more than ten years from Apple leaving Intel to M1 shipping. Model companies are now setting up custom chip teams, and there is still a distance to large-scale deployment. Tape-out, yield, compilers, drivers, and framework adaptation all take time. For developers, the more realistic near-term questions are supply and software support, not being able to replace GPUs tomorrow.

One reply says nobody likes paying tolls, so they simply buy the road. That gets at the motivation. Custom chips are like buying a road, not building one. Big companies can amortize long-term costs, while smaller companies pay tuition first. Chip teams, EDA, IP, and advanced-process capacity add up to a hefty bill.

Another reply frames the cycle more coldly: the market will become more diverse, but the process will be long. Right now everyone is using custom chips to lower costs within this cycle, and this cycle may already be set. The GPU effect won't appear immediately. The manufacturing layer may be the first to turn, via TSMC, and the cycle and cost of this layer alone are enough for the software market to go through two rounds. Hardware and software markets run on different cycles.

Some material puts this in the same week as Nvidia management changes, arguing that even when the incumbent's gross margins are fattest, it is also shifting gears, and labs have a window to push first-party chips. This judgment should be treated only as a netizen's observation; whether the timing holds is another matter. But it raises a variable worth tracking: when the largest supplier's rhythm changes, customers' sense of urgency about custom silicon will grow.

A more conservative judgment is: model companies building chips will not, in the short term, significantly lower API prices, nor will it immediately cost Nvidia orders. What is more likely is stratification. Super-large model vendors will use custom chips for inference optimization and bargaining leverage, while small and medium vendors will continue to buy ready-made compute. Chips are a capital-intensive business, and model companies first need to prove that their software stack can keep total cost of ownership under control.

For ordinary users, the felt impact will come very slowly. For developers, what is worth watching is inference cost, supply stability, and framework support. If custom chips only run in a company's own cloud, external developers may not be able to touch them. By the time they can, migration cost is another threshold.

Reading the replies is more useful than reading slogans.

发布时间: 2026-11-16 09:29