The buyer's market for inference
AMD spent the day rewriting its inference story. The company said Thursday it will acquire Taalas, a Toronto-based chip startup, for an undisclosed sum. Taalas' pitch is unusual enough to explain the interest: rather than building general-purpose accelerators that can run any model, it hardwires a single AI model directly into the silicon. That trades flexibility for efficiency, and it aims squarely at inference — the part of the AI business where models are actually served to users, day after day, at cost.
The strategic logic is easy to read. Training grabs headlines; inference grabs the recurring bill. AMD has spent two years trying to convince buyers it is a credible second source in AI hardware, and baking models into chips is a way to compete on something other than raw general-purpose throughput.
The market's reaction was the tell. Nvidia's stock went up anyway. That is the shape of this moment: the AI hardware pie is expanding fast enough that a rival's acquisition doesn't read as a threat.
Two playbooks, both winning
Which brings us to the quieter story underneath the day's news. Nvidia and Broadcom are chasing the same boom from opposite directions, and both are ahead. Nvidia sells general-purpose AI chips that anyone can buy and program — the approach most people picture when they think about AI hardware. Broadcom's path is less visible and, for now, no less lucrative: helping the giants design their own custom silicon. Every AMD-Taalas-style bet on specialization is a vote for the second model.
Anthropic joins the custom-chip club
The most striking new entrant is a company that doesn't make hardware at all. Anthropic is building its own AI chips.
The first sign came from job listings: the company has begun hiring Research Engineers for a "Chip Design RL" team based in San Francisco, per a report carried on MSN via Bing News. Anthropic has since confirmed it is assembling a team to co-design custom ASICs — application-specific integrated circuits — aimed at running Claude. Forbes frames the move as putting Anthropic alongside OpenAI, Meta and Google in the race for proprietary AI silicon.
That the team is called "Chip Design RL" is worth pausing on. It suggests the machine-learning methods are pointed at the chip design process itself, not just at what the finished chip will run.
Musk's very large building
Elon Musk is pooling money across his two biggest companies to get into chips directly. SpaceX and Tesla are investing $16.8 billion in a Texas semiconductor plant dubbed "Terafab," according to India Today, with coverage on MSN amplifying Musk's claim that it will be the world's largest building. Take the superlative as marketing; take the number as a signal of how much vertical integration is suddenly worth to companies that consume chips at scale.
Memory is the bottleneck — and the contest
TSMC, the world's largest contract chipmaker, raised its outlook for 2026, telling investors AI demand is still accelerating. Downstream, the squeeze is in memory. SK Hynix plans to spend roughly 54 trillion won on two new domestic fabs — reported as $38 billion by most outlets, $38.1 billion by Engadget — a response to AI memory running short.
Its lead may be narrowing. China's CXMT has pushed its DDR5 chips past the DDR5-8800 mark on an AMD platform, a demonstration Tom's Hardware frames as the homegrown memory champion closing the gap with SK hynix. A single benchmark isn't a supply chain, but it is a milestone that didn't exist a year ago.
And one from the other side of the lab
D-Wave says its $550 million January acquisition of Quantum Circuits is paying off: per a Forbes report by John Koetsier published August 7, the deal has made quantum error correction 10 times cheaper. Error correction is the tax that stands between quantum machines and useful work — cutting it by an order of magnitude is the kind of unglamorous progress that matters.