Every published Top AI Stories item tagged with Semiconductor Manufacturing, newest first.
Citing people briefed on classified intelligence, the New York Times reported that US agencies watched Alice Huang, a former top executive at the Singapore cloud company Megaspeed, for about a year over concerns that she was passing Beijing information from her access to Nvidia chief executive Jensen Huang, who is no relation. Megaspeed was already under scrutiny for allegedly helping Chinese firms reach restricted Nvidia chips. According to the Times, President Trump was briefed and chose not to act, partly over worries about Nvidia's sales and the trade truce with China. Nvidia said the meetings were ordinary customer contact and that regional customers receive "zero information relevant to national security." China's embassy called the accusations baseless.
Microsoft announced Maia 300, the third generation of its in-house AI chip, and said it will scale the Maia family to deployments of several gigawatts starting in 2027, making it the main system that serves AI answers across Microsoft products. The company expects Maia 300 to produce 40 percent more tokens per megawatt than leading GPU systems, and up to seven times the rack throughput of Maia 200 in a 96-chip configuration. It is built from chiplets on TSMC's N2P process, with eight HBM4 memory stacks. Maia 200 already serves Copilot and runs OpenAI models, with GitHub Copilot and Microsoft Foundry to follow by year end. The Verge framed the move as Microsoft becoming less dependent on Nvidia.
Micron chief executive Sanjay Mehrotra told investors that 75 percent of the company's memory output for 2027 is already accounted for, and that most sales talks now concern 2028. High-bandwidth memory for AI chips is outrunning demand for ordinary server memory. Samsung executive Kim Taewoo said it will take almost 30 percent of memory makers' wafer capacity in 2027, against 20 percent this year, Reuters reported. Micron's new clean rooms open in 2028 but ramp slowly, so the squeeze on PCs, phones and game consoles looks set to last.
Arm introduced the Mali G2-Ultra NX, which puts neural accelerators inside the graphics pipeline rather than beside it. Three techniques ship together: reconstructing high-resolution frames from cheaper renders, generating intermediate frames to reach 120 frames per second, and combining upscaling with denoising for ray tracing. Arm reports up to 24 percent higher benchmark performance than the prior generation and 14 percent on non-AI gaming. Tencent Games, Unity China and the studio behind Where Winds Meet are integrating it. No device timeline was given.
Prosecutors in Taoyuan searched Unimicron's headquarters and a manufacturing plant on August 28 and questioned 14 executives and staff, alleging the company shipped circuit boards made in mainland China back to Taiwan and relabeled them as Taiwanese. A department general manager was released on bail of 15 million New Taiwan dollars, about 474,000 US dollars. This is a separate matter from the server-smuggling charges filed in Taiwan a week earlier, and it points at a different weak spot — not where AI hardware ends up, but whether the paperwork saying where it was made can be trusted.
Hot Chips 2026 produced two answers to the same problem: inference is increasingly limited by memory rather than by arithmetic. Xcena and Samsung showed the MX1, an expansion card carrying up to 2 terabytes of DDR5 alongside 3,072 small RISC-V cores, so the work happens beside the memory instead of crossing a link that tops out near 128 gigabytes per second. Samsung separately presented a processing-in-memory version of the LPDDR5X memory used in phones, reaching 614 gigabytes per second inside the chip against 76.8 gigabytes per second out of it. The Samsung part is a prototype with no ship date.
Politico reported that the administration is preparing semiconductor tariffs that could extend well past chips themselves, to finished goods built with them — game consoles, and the servers that fill data centers among them. The Computer and Communications Industry Association estimates that version would cost the United States $90 billion a year in economic output and delay or cancel roughly 20 percent of data-center projects planned through 2030. Some development would move offshore instead. Trade groups have warned for months that taxing both chips and the products containing them would raise prices on phones, laptops and cars. Relief for AI firms is under discussion, but sources said it would be tied to foreign companies investing in American chip plants.
OpenAI presented Jalapeño at the Hot Chips conference and invited SemiAnalysis into its lab to run the InferenceX benchmark suite against it. The chip was designed with Broadcom from a blank sheet for language-model inference, went from initial hiring to tape-out in roughly 16 months, and uses HBM4 memory. SemiAnalysis measured more than 700 tokens per second per user on DeepSeek R1 at a concurrency of one, with no speculative decoding. Two caveats belong beside the headline, and SemiAnalysis states both: OpenAI supplied the numbers, and the fair comparison is Nvidia's shipping Rubin generation rather than Blackwell.
Prosecutors in Keelung charged nine people, one of whom is still at large, over Supermicro servers carrying Nvidia B300 GPUs that reached China through Indonesia, Japan and Hong Kong. Seventy-four got through and another fifty-six were seized in Taiwan before they moved. The accused include a distribution manager at Nvidia's Taiwan office, whom prosecutors call the key figure for authorizing release of the chips, two Supermicro sales managers and a distributor's chief executive. It is the island's first known criminal case over the black market in AI accelerators, and it locates the export-control failure inside the vendors rather than at the border.
Contract manufacturers that build AI servers for Microsoft, Google and Oracle have told customers to expect increases of more than 15 percent on many systems, taking effect on machines shipped early next year, Bloomberg reported. The rises cover the flagship Vera Rubin and Grace Blackwell generations and vary by chip generation and memory configuration. The cause is memory rather than silicon: Samsung, SK Hynix and Micron cannot keep pace with AI infrastructure demand, and Nvidia is passing the cost through instead of absorbing it. Nvidia did not respond to requests for comment.
Xiaomi's chip team announced three in-house processors on August 24: the three-nanometer Xring O3 for phones, the six-nanometer Xring O100, and the three-nanometer Xring D100 for self-driving cars. The O100 is the AI part — a neural processing unit that stacks memory vertically against the compute die to reach 1.22 terabits per second of bandwidth, built to run Xiaomi's own MiMo language model locally rather than in a data center. The D100 carries up to 160 gigabytes of unified memory and is specified for models of roughly 200 billion parameters. The O3 is in volume production; the other two are slated for market next year.
Micron unveiled Micron Research Labs, a $10 billion, decade-long research effort anchored on a flagship campus in Boise, Idaho, with groundbreaking expected in calendar 2027. It targets memory and compute architectures, packaging and future manufacturing more than ten years out rather than the next product cycle — the company calls it the first dedicated memory research hub of its kind in the United States. Stanford, the University of Texas at Austin, Applied Materials, Lam Research and imec are named partners. The commitment is separate from the more than $250 billion Micron has already pledged to US manufacturing and research.
Cerebras announced its CS-4 rack system, claiming more than 1,000 tokens per second on models above 10 trillion parameters and up to 30 times the speed of production graphics-processor systems. The headline part is what the chip is not: the **WSE-3 Turbo is the same 900,000-core, 5-nanometer wafer as the two-year-old WSE-3**, clocked from 1.4 gigahertz to 2.8 gigahertz rather than re-fabricated. OpenAI is using it for the Ultrafast tier, and AMD is pairing its graphics processors for prefill with Cerebras for token generation. Shipments start this quarter.
The transformer-specialized chip startup Etched raised $700 million, led by Jane Street, at a $21 billion valuation — roughly double the $10.3 billion it carried in July. Unusually for the quant fund, Jane Street is also a customer: it has installed Etched's first shipped cluster in its own data center and says it is "pleased with the early results." Etched now sells a two-part system, a prefill chip plus a cluster-scale memory design, rather than a single accelerator.
TSMC reported July revenue of $14.5 billion, up 44.7 percent from a year earlier and 5.6 percent from June, beating the monthly record it set in June. Revenue for the first seven months of the year now runs 37 percent above the same stretch of 2025, which the company attributes to shipments of chips built on its 2-nanometer process. TSMC makes the leading-edge silicon that nearly every frontier AI system depends on, so its monthly disclosure is one of the cleanest public reads on whether AI hardware spending is still accelerating.
Situational Awareness, the fund Leopold Aschenbrenner launched in 2024 on the thesis that AI progress was outrunning the market, has invested $400 million in Source Foundry, bringing its total in the company to $500 million. Source Foundry was started by Stanford alumni Abdulmalik Obaid and Joe Burg to make chip manufacturing faster and cheaper, and is valued at $5 billion. The bet follows a hard stretch for the fund, which sold most of its public portfolio to Citadel in late July as assets under management fell from $20 billion to $10 billion.
The two companies confirmed Grimes County, Texas — about an hour northwest of Houston — as the site for Terafab, a plant that would run logic, memory, packaging and testing under one roof rather than splitting them across suppliers. The first phase is $16.8 billion for a facility of more than 100 million square feet and at least 3,000 workers, making chips for Tesla's Optimus robots and Cybercab as well as SpaceX's planned space-based data centers. The caveat is in SpaceX's own filing, which describes the arrangement as a general framework with no binding commitments and no obligation for either side to keep going. No completion date has been given.
High Bandwidth Flash sits between high-bandwidth memory and solid-state storage, using NAND flash to hold far more data than HBM does at speeds much closer to it. The first specification, published through the Open Compute Project six months after the consortium formed, defines stacks up to 512 gigabytes with three performance grades running from roughly 0.4 to 3 terabytes a second. The target is inference, where model weights and context windows have outgrown the memory attached to each accelerator. Google and Tenstorrent are also in the consortium.
AMD used its Advancing AI event to launch Helios, its first full rack-scale system aimed squarely at Nvidia's grip on AI data centers. Each rack packs 72 of AMD's new Instinct MI455X GPUs with sixth-generation EPYC processors and Pensando networking, delivering up to 2.9 exaflops of inference performance. Microsoft will run Helios on Azure and Anthropic plans to install up to two gigawatts of the chips, with OpenAI and Meta also signed on. Systems ship in the second half of the year.
Etched, a startup building chips hardwired specifically for the transformer models behind today's chatbots, raised $300 million, led by Sequoia, more than doubling its valuation to $10.3 billion. Andreessen Horowitz, SK Hynix, and backers including Peter Thiel and Andrej Karpathy joined the round. Founded by three Harvard dropouts, Etched says it has already booked $1 billion in orders and is shipping its first systems — a bet that purpose-built silicon can undercut Nvidia's general-purpose GPUs on inference cost.
Surging AI data-center demand for memory is squeezing consumer electronics, as chip makers steer supply toward far more profitable AI servers. Counterpoint Research says mobile memory prices in the second quarter ran nearly double their late-2025 levels, helping drag global smartphone shipments down about 3 percent to their weakest second quarter in 13 years and ending nine straight quarters of growth. Budget phones are hit hardest — models priced under 150 dollars face what analysts call a near-permanent squeeze out of the market — with price-sensitive markets like India seeing the sharpest declines.
Reuters reports that DeepSeek has spent about a year secretly recruiting chip engineers and courting manufacturing partners to build an in-house processor for AI inference — the stage where a trained model answers user queries. The goal is to cut its dependence on Nvidia and Huawei as US export controls keep tightening. It follows OpenAI's Broadcom-built Jalapeno inference chip and Anthropic's own reported chip ambitions — a sign that frontier labs increasingly want to own their silicon.
SambaNova, which designs chips tuned to run large models efficiently, raised $1 billion in a round led by General Atlantic that values it at $11 billion — just five months after its last mega-round. Backers include Intel, BlackRock, and T. Rowe Price. The company pitches itself as an inference-focused alternative to Nvidia, and its CEO signaled that a public listing, not a sale, is the likely path.
Etched, the startup building chips specialized purely for AI inference rather than general-purpose GPUs, said it has booked $1 billion in contracts for its frontier inference clusters and carries a $5 billion valuation. Backed by Jane Street, Hudson River Trading, and angels including Andrej Karpathy and Geoffrey Hinton, the Harvard-dropout-founded company is chasing the inference bottleneck that dominates AI costs at scale — where it takes on Nvidia, Cerebras, and Groq.
President Lee Jae-myung unveiled three "mega projects" totaling roughly $880 billion over ten years — about 5 percent of the country's GDP. The package pours $518 billion into new Samsung and SK Hynix memory fabs to ease the AI-driven shortage some call "RAMageddon," funds $356 billion of AI data centers, and aims to lift South Korea's share of the humanoid-robot market from 1 percent to 20 percent, with Hyundai's Boston Dynamics mass-producing Atlas robots by 2028.
AI chipmaker Groq confirmed a $650 million funding round, led by the investment firm Disruptive, as it rebuilds after Nvidia's roughly $20 billion deal late last year to license Groq's technology and hire away its founder and much of its top engineering team. Groq is now doubling down on its "neocloud" business — renting out fast inference capacity — and restocking its executive bench. The startup last carried a valuation near $6.9 billion and did not disclose a new one.
Qualcomm is reportedly negotiating to acquire Tenstorrent, the AI-chip startup led by legendary designer Jim Keller, in a deal valued between $8 billion and $10 billion. Tenstorrent builds RISC-V-based processors it claims beat general-purpose GPUs on specific AI workloads, so the purchase would hand Qualcomm a credible data-center play against Nvidia and a path beyond smartphone chips. The talks are unconfirmed and could still fall apart, but they show how aggressively mobile-chip makers are now chasing AI silicon.
Amazon is in talks to sell its in-house Trainium chips directly to outside data centers for the first time, breaking from years of reserving them for AWS cloud customers. AI chief Peter DeSantis confirmed the discussions in Paris but named no buyers, pointing to rising demand — especially in Europe — to keep AI compute under local control. CEO Andy Jassy has called the homegrown chip business an opportunity worth 50 billion dollars a year; pulling it off would put Amazon in direct hardware competition with Nvidia for the first time.
Commerce Secretary Howard Lutnick told ASML executives that Washington believes one of the company's extreme ultraviolet lithography (EUV) machines — the kind banned from export to China — may have ended up there, Bloomberg reported. ASML pushed back hard, circulating a document titled "No indication of any ASML EUV system in China" and noting it has never shipped an EUV system, component, or module to the country. No evidence of an actual transfer has been made public, and the systems weigh roughly 180 metric tons, making a quiet move unlikely.
Nvidia sold $25 billion of investment-grade notes across seven tranches maturing as far out as 2056 — its first corporate bond deal since a $5 billion raise in 2021. Demand reached more than $85 billion, over three times the offering, and the company upsized the sale from an initial $20 billion. For a firm sitting on a roughly $5 trillion market value and ample cash, tapping the debt market signals just how much it plans to spend building out AI data-center capacity.
Enflame, a Shanghai designer of AI training and inference chips, has won approval for an initial public offering on the STAR Market, where it plans to raise about 6 billion yuan — roughly $888 million — to fund future chip generations. Tencent owns about 20 percent of Enflame and supplies the bulk of its revenue, making the listing a test of how far Chinese buyers will bankroll a homegrown answer to NVIDIA, which still holds around 70 percent of the domestic market. Enflame is the last of China's 'four little dragons' of AI chips to head public.
Nvidia told Chinese cloud customers that its new Vera data-center CPU is now open for orders and could ship as early as August, according to Reuters. The pitch is a workaround: after US export curbs choked Nvidia's GPU sales in China, the company is leaning on standalone CPUs — which face looser restrictions — to defend the market. Nvidia's finance chief has said Vera CPU revenue could approach $20 billion in the coming fiscal year.
Amazon made its Graviton5 processor generally available, pitching it as purpose-built for the real-time reasoning, code generation, and multi-step orchestration that AI agents demand. AWS says the chip delivers up to 25 percent better compute performance than the prior generation, with 192 cores per processor and 33 percent lower inter-core latency. Uber and Snowflake are among the early adopters, joining more than 120,000 customers already running on Graviton. The launch underscores how the cloud giants are racing to design their own silicon and cut their dependence on Nvidia for the inference work that agents run around the clock.
Jensen Huang and Hyundai chair Chung Euisun used a meeting in Seoul to widen their partnership from research prototypes toward factory-ready robotics, spanning mobility, manufacturing, and autonomous driving. Hyundai will stand up an AI supercomputer built on 50,000 Nvidia Blackwell GPUs to train models for self-driving and smart factories, and the two companies will work with the Korean government on new AI technology centers. Huang called Hyundai one of the best robotics companies in the world. The deal is one of the largest national commitments yet to physical AI — the effort to move machine intelligence off the screen and into machines that move.
Bloomberg reports that Beijing is preparing to spend roughly 2 trillion yuan, about $295 billion, over five years on a network of interconnected AI computing hubs run mostly by state firms like China Mobile and China Telecom. The blueprint, drafted by agencies including the National Development and Reform Commission, would source at least 80 percent of the hardware, including AI chips, from domestic suppliers such as Huawei — effectively designing Nvidia and AMD out of the country's largest buildout. It is the clearest signal yet that China intends to win the AI compute race on homegrown silicon.
At Computex, Intel detailed Crescent Island, a data-center GPU built on its Arc Xe3P architecture and aimed at AI inference. Partner cards can carry up to 480 gigabytes of LPDDR5X memory — more than Nvidia's and AMD's flagships — while skipping the scarce, expensive high-bandwidth memory those chips rely on, which Intel says makes the card cheaper to produce and run on a 350-watt air cooler. Intel is targeting a second-half 2026 launch as it tries to chip away at Nvidia's data-center lead.
Korean chip startup Xcena closed a $135 million Series B at a $570 million valuation on Friday, co-led by Atinum and IMM Investment with Corstone Asia, SBI Investment, and Mirae Asset Capital joining. Its MX1 chip embeds compute directly inside DRAM modules over the CXL interconnect, aiming to cut the round-trips between GPU and memory that increasingly bottleneck inference. The founders are former Samsung and SK Hynix veterans, and Xcena says the design could let workloads that currently need 10 servers fit on one — though it has not named a launch customer.
Groq is raising 650 million dollars to refocus the company on its inference neocloud — the on-demand cloud platform powered by Groq's own AI chips — after a December 2025 arrangement saw Nvidia pay 20 billion dollars for senior Groq engineering talent and a hardware license. Existing investors are leading the round, with Disruptive and Infinitium committed to fill any unsubscribed shares. Adam Winter is now interim CEO and Matt Eng interim CFO. The company argues inference is a much bigger market than training right now, and the wedge for the smaller team that remains.
At the AI Now Summit, Mistral named Airbus (commercial aircraft, helicopters, defense, and space), BMW Group (the carmaker's Large Industry Model initiative for multimodal reasoning on engineering data), and ASML (semiconductor parts design and surrogate modeling) as production customers for its new Mistral Physics AI line. The lab also confirmed a new 10 megawatt data center in Les Ulis south of Paris, dedicated to inference and scheduled to open in the third quarter of 2026, alongside an existing 40 megawatt Paris site and a planned Swedish facility — Mistral's answer to the inference supply-chain risk that has loomed over European AI customers.
At the groundbreaking for a new Taipei headquarters, Jensen Huang announced that Nvidia will spend $150 billion a year in Taiwan, up from $10 to $15 billion annually four to five years ago. The headquarters will employ 4,000 people and target a 2030 opening, anchoring Nvidia closer to TSMC, which fabricates its chips, and Foxconn, which assembles them into server racks. The pledge lands as Washington has spent two years pushing chipmakers to build US capacity, and it cements Taiwan rather than the United States as the structural center of advanced AI manufacturing for at least the rest of the decade.
Mistral acquired Vienna-based Emmi AI and used the deal to launch Physics AI, a class of data-driven models that learn from physics solver outputs and predict the behavior of physical systems in seconds on a single GPU, replacing simulations that traditionally take hours or weeks. Mistral named Airbus, ASML, Safran, and Siemens Energy as launch partners, targeting aerospace, semiconductors, energy, and industrial equipment. The offering pitches three concrete use cases: exploring thousands of design variants for new products, optimizing factory tooling, and running real-time digital twins on live sensor data — engineering acceleration with named industrial deployments rather than future-tense lab promises.
In this week's Stratechery analysis, Ben Thompson argues that Nvidia's revised reporting taxonomy — splitting hyperscaler revenue from everyone else — reveals the contours of an emerging commoditization fight at the top of the stack. Nvidia is "fighting commoditization" with its largest customers, where hyperscalers increasingly design their own silicon, while it "runs the whole stack" for AI clouds, sovereigns, and enterprises. The reframe lands one week after Nvidia's record $81.6 billion quarter on May 20, when hyperscaler revenue held at roughly half of the $75.2 billion data-center total.
At a press event in Taipei this week, NVIDIA CEO Jensen Huang told reporters the company has *"largely conceded"* China's AI accelerator market to Huawei, with NVIDIA's share now near zero after the US H200 export-clearance stalemate dragged into a second month. Huawei expects roughly $12 billion in 2026 revenue from its Ascend line — up from $7.5 billion in 2025 — on orders already placed by Alibaba, ByteDance, and Tencent, all of which deployed DeepSeek V4 services within hours of the model's Ascend-optimized release in April. Huang said he still expects Beijing to eventually allow H200 imports, but for now the homegrown stack is shipping while NVIDIA's clearance letters sit in customs.
The Department of Commerce signed letters of intent on Thursday to provide $2.013 billion in federal incentives to nine quantum companies — IBM alone receives $1 billion, GlobalFoundries gets $375 million, and seven computing firms (D-Wave, Rigetti, Quantinuum, PsiQuantum, Atom Computing, Infleqtion, and Diraq) get between $38 million and $100 million each. Unlike traditional CHIPS Act grants, the department takes a **minority, non-controlling equity stake** in each company "to enhance the return for the U.S. taxpayer," mirroring the structure of last year's Intel deal. Secretary Howard Lutnick framed the package as building "thousands of high-paying American jobs while advancing American quantum capabilities."
Nvidia reported $81.6 billion in revenue for the quarter ending April 26, a 20% sequential jump, with data-center revenue at a record $75.2 billion. The earnings disclosure also surfaced $43 billion in non-marketable startup equity — nearly double the prior quarter — including a previously-undisclosed []0 billion commitment to OpenAI. On the call, Jensen Huang positioned the new Vera CPU as a "brand new toolIds: 00 billion total addressable market" built for autonomous-acting agents.
Per SpaceX's S-1 filing, Anthropic will pay xAI $1.25 billion per month for 300 megawatts of capacity at the Colossus 1 data center in Memphis, running through May 2029 — a contract worth over $40 billion. xAI burned $6.4 billion last year and Grok usage has dropped sharply, leaving the lab with excess capacity it is now monetizing by selling to a direct competitor. The arrangement underscores how blurred the line between rival AI labs and their cloud-supplier parents has become.
Mistral closed an acquisition of Linz-based Emmi AI, an Austrian engineering startup building physics-based simulation models for aerospace, automotive, semiconductors, and energy. Roughly 30 researchers and engineers join Mistral's team, and Linz becomes an official Mistral office. The combined pitch: real-time simulations and digital twins layered onto Mistral's platform to court industrial buyers in high-stakes manufacturing sectors. Terms were not disclosed; the deal closes in May 2026.
Cerebras Systems priced its IPO at $185 per share — above its $115 to $160 range — and raised roughly $5.5 billion before the stock more than doubled on debut Thursday, closing at $311 for a $66 billion market cap. The S-1 names OpenAI, Group 42, Saudi Arabia's MBZUAI, and Amazon Web Services as top customers, and Cerebras swung to profitability on $510 million of 2025 revenue. The pop validates the thesis that frontier labs will pay a premium for Nvidia alternatives if inference economics work — and gives Cerebras a public-market currency to chase Nvidia's data-center share more aggressively.
In this week's Stratechery, Ben Thompson argues that OpenAI's and Anthropic's new "deployment company" structures (engineers embedded in enterprises to ship AI systems) mirror the 1970s mainframe wave: executive-mandated automation that eliminates jobs rather than employee-driven productivity tools. He pairs this with an Apple–Intel angle — TSMC's AI-chip prioritization is forcing Apple toward Intel as a secondary fab, giving Intel "the single most important thing it needs" to compete: a marquee customer eager to reduce TSMC dependency.
In this week's Stratechery analysis, Ben Thompson argues that AI compute is bifurcating into three distinct workload categories that need fundamentally different hardware. Training keeps high-bandwidth GPUs (NVIDIA's lock-in); answer inference rewards token speed (Cerebras's WSE-3 packs 44 gigabytes of on-chip SRAM at 21 petabytes per second of bandwidth versus NVIDIA H100's 80 gigabytes of HBM at 3.35 terabytes per second); agentic inference, where humans aren't in the loop, mostly cares about memory capacity and cost-per-token at scale. Thompson treats Cerebras's revised IPO pricing of $150 to $160 per share (up from $115 to $125) and Anthropic's lease of 220,000 NVIDIA GPUs at SpaceX's Colossus 1 (the 300-megawatt Memphis data center) as concrete evidence that buyers are now sorting their compute spend by workload category rather than picking a single vendor.
Nvidia has now committed more than $40 billion to equity investments in AI companies in 2026, including a single $30 billion stake in OpenAI, up to $3.2 billion in glassmaker Corning, and up to $2.1 billion in data-center operator IREN. Wedbush analyst Matthew Bryson described the pattern as "squarely circular" — money cycling between chip vendor, model customer, and infrastructure provider. The chipmaker has also closed roughly two dozen private startup rounds plus 67 venture deals across 2025, intensifying scrutiny of how concentrated the AI capital stack has become around a single supplier.
NVIDIA Research published Nemotron Elastic, a post-training method that embeds 30, 23, and
12 billion-parameter nested reasoning models inside a single 30 billion-parameter parent —
extractable via zero-shot slicing without further fine-tuning. The recipe achieves a 360-times
token reduction over pretraining from scratch, and the 30 billion checkpoint compresses to 18.7
gigabytes under NVFP4 quantization. The 23-to-30 billion configuration advances the
accuracy-and-latency Pareto frontier with up to 16 percent higher accuracy and 1.9-times lower
latency than the default Nemotron Nano v3 budget control. All three precision variants (BF16,
FP8, NVFP4) are available on Hugging Face under nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B.
San Francisco lab Zyphra released ZAYA1-8B, an 8.4 billion-parameter mixture-of-experts model with **only 760 million active parameters at inference. On math benchmarks the model matches DeepSeek-R1 (89.1 on AIME 2026, 71.6 on HMMT)** and stays competitive with Claude Sonnet 4.5 on reasoning, despite running roughly an order of magnitude smaller. Most notably, the model was trained on a **1,024-node AMD Instinct MI300X cluster with AMD Pensando Pollara networking** — the first frontier-quality math model trained outside the NVIDIA CUDA stack. Weights are on Hugging Face under Apache 2.0.
Cerebras Systems filed to sell 28 million shares priced between $115 and $125 per share, targeting $3.5 billion in proceeds at a $26.6 billion valuation — the largest US tech IPO of 2026 so far. OpenAI is one of the chipmaker's largest customers under a multi-year contract worth more than $10 billion signed in January, and holds a $1 billion secured loan plus warrants for over 33 million shares, potentially making OpenAI a major shareholder post-listing. The offering puts Cerebras' Wafer-Scale Engine 3 against Nvidia in a public-market test of GPU pricing power and frontier-lab compute lock-in.
The x86 Ecosystem Advisory Group has published the **AI Compute Extensions** whitepaper, jointly authored by AMD and Intel engineers. ACE adds two-dimensional tile registers and outer-product matrix-multiply instructions on top of AVX10, with a claimed 16-times compute-density improvement on matrix workloads. Unlike Intel's Xeon-only AMX, ACE targets both client and server silicon, and software enablement is already in flight for PyTorch, NumPy, and TensorFlow. Neither vendor has yet named which CPU generation will ship the new instructions.