AI chips coverage across GPUs, accelerators, custom silicon, memory shortages, foundries, export controls, and the semiconductor supply chain behind AI compute.
Stories
997
Latest source update
July 15, 2026
Coverage
Live
Topic brief
What to know about AI Chips
Brief updated Jul 11, 2026
AI chips are the specialized processors and memory that make modern machine learning economically possible. The category spans training and inference accelerators (Nvidia and AMD GPUs, Google TPUs, and custom silicon such as Amazon Trainium and Meta MTIA), a fast-growing field of inference-focused startups (Cerebras, SambaNova, Groq, d-Matrix, and others), the high-bandwidth memory (HBM) and advanced DRAM that feed those accelerators, and the foundries and packaging that manufacture them. Because compute is the binding constraint on AI progress, the chip layer sits underneath every model, product, and data center in the industry.
For practitioners, chips determine what is possible and what it costs. ML engineers and infrastructure teams care about accelerator availability, memory bandwidth, interconnect, and the price-performance of inference, which increasingly dominates real workloads. Finance and procurement teams track rack costs, lead times, and the memory market, because an HBM shortage or a delayed rack system can reshape a build-out. Startups and researchers watch novel architectures and numeric formats that promise lower cost and power. And almost everyone is exposed to the supply chain, since a handful of foundries and memory makers underpin the entire market.
The market structure has three defining features. First, Nvidia remains dominant but is no longer unchallenged, facing custom silicon from its largest customers, a crowd of inference-chip startups, and its own manufacturing and roadmap constraints. Second, memory has become the swing factor, as AI demand drives a supercycle in HBM and DRAM that lifts prices across consumer and data-center hardware alike. Third, chips are deeply geopolitical, with export controls, national industrial policy, and a push for domestic and sovereign manufacturing shaping who can buy and build what. Following AI chips means tracking accelerators, memory, foundries, and policy together.
What changed recently
The chip market keeps booming and repricing risk at the same time. Nvidia shares have fallen roughly 16% from their May peak, erasing about $1 trillion in value, and the company pushed its Kyber NVL144 rack system out to 2028 over a hard-to-manufacture midplane, even as analyst estimates put a single Rubin Ultra rack near $21 million. Memory remains the clearest demand signal: Samsung reported record quarterly profit on AI memory, and Micron lifted planned U.S. investment above $250 billion through 2035 while continuing a $9.3 billion expansion of its Hiroshima, Japan plant for HBM. Foundry capacity is racing to keep pace, with Japan's Rapidus pursuing 2nm production for 2027. On the demand side, Meta said it plans to put its in-house MTIA chip into production in September 2026 as it works toward roughly doubling data-center computing capacity to 14 gigawatts by 2027, underscoring how custom silicon is becoming core to hyperscaler capacity plans rather than a side bet.
The challenger and geopolitics threads are intensifying together. SambaNova raised $1 billion at an $11 billion valuation and Cerebras is scaling toward 200 MW of European inference capacity, both chasing the inference workloads that increasingly dominate deployment, while China's Biren raised roughly $892 million to scale GPU production and DeepSeek is reportedly building its own inference chip to cut reliance on Nvidia and Huawei. China is also weighing limited Nvidia H200 access, possibly fewer than 200,000 chips, for top domestic firms including Alibaba, ByteDance, and DeepSeek, and on July 10 temporarily halted helium exports, a gas used in semiconductor manufacturing, as Middle East conflict strained global supply, adding a fresh supply-chain risk on top of ongoing HBM and packaging constraints.
What to watch
Watch the timing and financing milestones that will set capacity for the next two years. Nvidia's Kyber NVL144 rack is now slated for 2028, so buyers tied to Rubin Ultra roadmaps should plan for later or reallocated capacity, and the roughly $21 million estimated cost of a Rubin Ultra rack shows where budgets are heading. Meta's MTIA chip is set to enter production in September 2026 as part of a push toward 14 gigawatts of data-center capacity by 2027, a concrete test of whether hyperscaler custom silicon can scale on schedule. Cerebras aims to bring first European capacity online by end-2026 and reach 200 MW by end-2027, Micron's more than $250 billion U.S. plan runs through 2035 alongside its Japan HBM expansion, and Rapidus targets 2nm mass production in 2027, all long-dated bets. On the geopolitical side, watch whether China's temporary helium export halt disrupts chip manufacturing beyond the Middle East conflict that triggered it, whether Beijing finalizes its reportedly limited Nvidia H200 allocation for firms like Alibaba, ByteDance, and DeepSeek, and whether DeepSeek's in-house inference chip effort or Biren's newly funded GPU scale-up narrows China's reliance on foreign silicon. Underneath it all, whether the memory supercycle keeps lifting prices is the variable that touches every buyer.
Comparison
segment
key players
recent signal
Incumbent GPU makers
Nvidia, AMD
Nvidia valuation fell about 16% from its May peak and delayed Kyber NVL144 to 2028; AMD shipped its Ryzen AI Halo appliance
Custom silicon (hyperscaler/ASIC)
Meta, Broadcom, Apple
Meta targets September 2026 production for its in-house MTIA chip as part of a 14GW capacity push; Broadcom extended its custom-ASIC supply deal with Apple through 2031
Inference-chip challengers
Cerebras, SambaNova, Tensordyne
SambaNova raised $1B at $11B valuation; Cerebras is scaling to 200MW in Europe and expanding CS-3 output with Flex; Tensordyne shipped a logarithmic-math chip on TSMC 3nm
Memory (HBM/DRAM)
Samsung, Micron, SK Hynix
Samsung posted record profits on AI memory; Micron raised U.S. investment above $250B through 2035 and began a $9.3B HBM expansion in Japan
Foundry and process
TSMC, Rapidus
TSMC's advanced nodes and packaging underpin new accelerators; Rapidus targets 2nm mass production in 2027
China self-sufficiency
Biren, DeepSeek, Huawei
Biren raised about $892M for GPU production; DeepSeek is reportedly building its own inference chip; Huawei plans a Q4 2026 South Korea launch for Ascend 950 chips
Frequently asked questions
Why are memory chips such a big deal in AI right now?+
Modern accelerators need enormous memory bandwidth, and high-bandwidth memory (HBM) plus advanced DRAM have become both the bottleneck and the profit center of the AI hardware market. AI demand has driven a supercycle: Samsung reported record quarterly profit driven by AI memory, and Micron raised planned U.S. investment above $250 billion through 2035 while also advancing a $9.3 billion HBM expansion in Japan. The knock-on effect is higher prices for data-center racks and even consumer electronics, so memory supply is now something every AI buyer has to plan around.
Is Nvidia still dominant in AI chips?+
Yes, but less unquestioned than before. Nvidia still anchors the market with its GPU and rack roadmap, but its shares fell about 16% from a May 2026 peak, erasing roughly $1 trillion in value, it delayed its Kyber NVL144 rack system to 2028 over a hard-to-manufacture component, and analyst estimates put a single Rubin Ultra rack near $21 million. At the same time, custom silicon from its largest customers and a wave of inference-chip startups are chipping at specific workloads. Nvidia remains the default, but buyers now have more alternatives and more reason to scrutinize cost and timing.
What are the main alternatives to Nvidia GPUs?+
Several categories. Hyperscalers build custom silicon such as Meta's MTIA, slated for production in September 2026, and Amazon Trainium, often with Broadcom, which also extended its custom-ASIC supply deal with Apple through 2031. Inference-focused startups including Cerebras, SambaNova, Groq, d-Matrix, and Tensordyne target the serving workloads that increasingly dominate production, sometimes with novel numeric formats. AMD offers competing GPUs and edge appliances. The common theme is that inference, not training, is where most of the new competition is concentrated, because that is where volume and cost now sit.
How is China responding to chip export controls?+
By pushing self-sufficiency across the stack while also managing its own leverage. Chinese lab DeepSeek is reportedly building its own AI inference chip, GPU maker Biren raised roughly $892 million to scale production, and Huawei plans a fourth-quarter 2026 South Korea launch for its Ascend 950 chips. China is also reportedly weighing limited domestic access to Nvidia's H200 for firms including Alibaba, ByteDance, and DeepSeek, and briefly halted helium exports, a gas used in chip manufacturing, as Middle East conflict strained global supply. The direction is a parallel Chinese chip ecosystem developing under export restrictions, though it still depends on foundry and memory partners.
Why does inference get so much attention versus training?+
Because inference is where most real-world cost and volume live once a model is deployed, and it has different hardware needs than training. That is why so much new investment targets inference: SambaNova raised $1 billion for AI inference, Cerebras is building large inference-capable capacity, and startups like Tensordyne use logarithmic math to make matrix multiplication cheaper. For practitioners, optimizing inference price-performance often matters more day to day than peak training throughput.
What role do foundries like TSMC play?+
Foundries manufacture the actual chips, and advanced process nodes plus packaging are gating factors for the whole industry. TSMC's leading-edge nodes and packaging underpin the next generation of accelerators, and Japan's Rapidus is pursuing 2nm mass production for 2027 as a supply-chain alternative. Even well-funded chip designs depend on foundry capacity and advanced packaging, which reporting flags as a persistent bottleneck. Foundry access is therefore a strategic constraint, not just a manufacturing detail.
