Musk's Colossus 1 AI supercomputer's inefficient mixed-architecture design couldn't be used to train Grok, so Anthropic's using it for inference instead
And now Anthropic is getting in on the act, and for much the same reasons. Anthropic told Business Insider that it was co-designing the chips so that they would allow Claude to run faster and more efficiently at the scale its customers need. Indeed, Anthropic has seen explosive growth in the past year, seeing huge expansion in the consumer space and taking on significant government contracts — not to mention agentic AI requiring far more tokens than traditional single-prompt interactions.
Anthropic hasn't revealed which firm it's working with on the design and development. Although Broadcom and Marvell are the two largest companies in the ASIC co-design market, representing some 95% of it, The Information reported last month that Anthropic was in talks with Samsung for manufacturing.
Designing, packaging, and manufacturing your own custom ASIC for AI inferencing isn't cheap, and it isn't easy. Alongside developing the hardware, you need the software stack to utilize it, and you want models that are optimized to run on it to make the most of its potential advantages. That makes it more worthwhile for most companies to simply use other firms' hardware and more general-purpose GPUs. But for major AI companies with enough money to burn and the ability to scale up to maximize efficiency gains, it's well worth the investment.
But the AI developers aren't the only ones who benefit. Nvidia has been one of the few companies to make enormous profits from the AI boom, while the likes of Meta, Google, Microsoft, and OpenAI are all losing enormous sums of money on their AI efforts. There are very real winners from the custom ASIC design and build market, too: TSMC, Broadcom, and Marvell.
Broadcom has been Google's co-design partner for years, and was also recently tapped to help build OpenAI's inferencing chips. It also works with Meta on its MTIA design, as well as holding contracts for other custom ASIC designs with ByteDance and Fujitsu. It also produces strong interconnect and networking hardware, which allows it to offer customers a more complete solution. It claims to have a $73 billion backlog of orders to work through, and expects to generate over $100 billion in annual AI chip revenue by the end of 2027.
Anthropic warns Claude AI is building itself faster than expected, calls for option to halt frontier development
AI is starting to out-design chip engineers in narrow areas as LLMs accelerate software chip design tool development
Marvell holds massive contracts with Amazon for its Trainium chips and Microsoft for Maia, and is expected to make upwards of $11 billion for these co-design jobs in 2026.
If Samsung ends up as Anthropic's partner, it would be a relatively small player in this particular space, but it would bring enormous manufacturing and chip design expertise to the table, as well as access to the all-important memory that is in such short global supply.
The biggest winner of all these initiatives, though, is arguably TSMC. The Taiwanese company produces the majority of the world's cutting-edge silicon and is involved in the production of almost all the chips discussed here. They need TSMC's CoWoS advanced packaging technologies for integration with HBM. It also handles much of the packaging of Nvidia's and AMD GPUs, as well as producing much of the underlying wafers.
Anthropic joining the custom ASIC race is hardly surprising and further cements the future we seem to be barrelling towards, which is each of the hyperscaler AI companies looking to handle as much of their inferencing with custom hardware as possible. It's more efficient, easier to control for features and specifications, and easier to scale up when optimized for internal models. From a Chinese perspective, it's also easier to avoid problems caused by international trade blockades and tariffs.
They'll likely never become 100% reliant on their own chips — there is just too much AI demand to scale into for that to happen, and Nvidia has been ruthlessly dominating access to the supply chain. But every new chip installed is an Nvidia GPU that won't be used for the same purpose, which may help reduce the stranglehold Nvidia has on the industry. Not for training, though. That's likely to remain Nvidia's biggest appeal for some time to come.
Jon Martindale is a contributing writer for Tom's Hardware. For the past 20 years, he's been writing about PC components, emerging technologies, and the latest software advances. His deep and broad journalistic experience gives him unique insights into the most exciting technology trends of today and tomorrow. ","collapsible":{"enabled":true,"maxHeight":250,"readMoreText":"Read more","readLessText":"Read less"}}), "https://slice.vanilla.futurecdn.net/13-4-25/js/authorBio.js"); } else { console.error('%c FTE ','background: #9306F9; color: #ffffff','no lazy slice hydration function available'); } Jon Martindale Freelance Writer Jon Martindale is a contributing writer for Tom's Hardware. For the past 20 years, he's been writing about PC components, emerging technologies, and the latest software advances. His deep and broad journalistic experience gives him unique insights into the most exciting technology trends of today and tomorrow.
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/tech-industry/SPONSORED_LINK_URL
- https://www.tomshardware.com/tech-industry/anthropic-to-build-its-own-co-designed-custom-ai-accelerator-for-inferencing-workloads-samsung-reported-to-be-partnering-with-the-claude-ai-maker-for-manufacturing#main
- https://www.tomshardware.com/my-account
- MSI's magnificent 32-inch 4K OLED gaming monitor returns to its lowest-ever price of $599 — save $130 on the MAG 321UPXB, the perfect 240Hz upgrade
- Save $1,400 on this RTX 5090 gaming PC from HP with 64GB DDR5 right now; RAM and GPU cost $5k alone — $5,599 sale price for top-spec Omen Max 45L rig with 24-co
- Texas slams on the brakes for 1,800 data centers, power grid requirements are five times higher than peak record demand — 474 gigawatts of power requests are no
- Microsoft wants the next-gen Xbox Helix to play every Xbox game ever made as it urges publishers to opt in — New report also claims Xbox 360 games coming to PC
- PC cooling outfit Arctic reverses tariff-era price hikes after US government refund — lowers prices across lineup, including coolers and case fans
Informational only. No financial advice. Do your own research.