Tenstorrent raises $1.5 billion, SpaceX earns $7.6 billion renting AI compute, and more - Silicon Brief #93

In this edition, covering 19 to 24 October:

• Tenstorrent closed a $1.5 billion Series E at a $6.2 billion valuation to deploy its Galaxy servers and develop its RISC-V CPU and AI core designs.
• SpaceX's AI revenue rose more than ninefold to $7.6 billion in the third quarter, mainly from renting computing capacity to Google, Anthropic and Reflection AI.
• ByteDance, Alibaba and Tencent have ordered more than 3 million Nvidia RTX PRO 5500 GPUs worth over $35 billion, people familiar with the orders say.

From our lists

The ten most valuable public AI chip companies by market value, from NVIDIA, Broadcom and AMD to Cerebras, Cambricon and Hygon.

10 most valuable public AI chip companies

Tenstorrent

Tenstorrent closed a $1.5 billion Series E at a $6.2 billion post money valuation, led by Trace Capital and Innovation X. The money goes to deploying its Galaxy servers at scale and to the RISC-V CPU and AI core designs it licenses to chip designers such as Samsung and LG. Series E

Japan's NEDO selected a joint proposal from Preferred Networks and Tenstorrent for a five year project to develop AI processors that combine photonics and electronics, including high bandwidth AI chiplets, open CPU chiplets with confidential computing and advanced packaging. NEDO

Hyperscalers and their spending

SpaceX reported third quarter revenue of $14.58 billion, with AI revenue up more than ninefold to $7.6 billion, mainly from renting computing capacity to companies including Google, Anthropic and Reflection AI. It spent $16.4 billion on AI infrastructure and plans to add 10 gigawatts of compute by 2027. In the same filing it disclosed that Mississippi's environmental regulator is seeking up to about $6.1 million from its subsidiary MZX Tech for starting construction at three data center sites before getting stormwater permits. SpaceX results · Mississippi

Meta's third quarter revenue rose 26% to $64.6 billion, and it plans publicly listed nonvoting Class C shares in January 2027. Meta says the first phase of its Prometheus supercluster in New Albany, Ohio, is complete, with hundreds of thousands of GPUs, and it plans to grow the site to about 1 million GPUs. Meta results · Prometheus

Amazon CEO Andy Jassy says AWS capacity is largely sold out through 2027, with significant commitments into 2028, and Amazon plans about $220 billion in capital spending in 2026. Amazon also plans more than €16 billion of investment in Germany in 2027, including €4.8 billion to expand its Frankfurt AWS region. Alphabet CEO Sundar Pichai says Google will keep giving computing resources to frontier AI research even when that reduces short term cloud revenue. AWS capacity · Amazon in Germany · Pichai

S&P Global's head of digital infrastructure, Mark Habib, said the largest tech companies likely carry more AI linked liabilities than disclosed, through financing structures that guarantee data center builds or smaller AI companies, though S&P sees no need to change ratings now. S&P expects the top six US hyperscalers to spend more than $7 trillion on AI related capital spending from 2025 to 2030. S&P Global

Cloud platforms

Google Cloud is rolling out OpenAI's GPT Astra, GPT 6.1 Sol and GPT 6 Luna models in private preview through Model Garden for select enterprise customers, and made a deployment of Claude managed by Anthropic generally available. Google Cloud will make bare metal A5X instances on NVIDIA's Vera Rubin NVL72 generally available next month, with new local regions for EMEA data residency, and launched Agent Compute Engine for running AI agents in hardware isolated microVMs. OpenAI models · Claude · A5X · Agent Compute Engine

Databricks made AI Runtime, its managed on demand GPU infrastructure, generally available. Amazon will open source TorchNeuron so PyTorch code runs on Trainium2 and Trainium3 with one configuration change, and Deepgram says running its Nova-3 speech model on Trainium2 processes 11 to 70 percent more audio per dollar. Databricks · TorchNeuron · Deepgram

Nvidia

NVIDIA plans to expand manufacturing capacity with Amkor, IBIDEN, King Yuan Electronics, Samsung Electro-Mechanics, SPIL and Unimicron, including a 2.5 times increase in substrate capacity over three years and new US testing facilities backed by a $1.1 billion advance to KYEC. Samsung Electro-Mechanics and NVIDIA will jointly invest in dedicated ABF substrate plants in Sejong and Busan. Supply chain · Samsung Electro-Mechanics

ByteDance, Alibaba and Tencent have ordered more than 3 million RTX PRO 5500 workstation GPUs worth over $35 billion, according to people familiar with the orders; the chips comply with US export controls, and Nvidia has warned it may not be able to supply the quantities. Cloud startup Modal will use Nvidia's Vera CPUs to host AI agents after early tests showed Vera doubled the work agents could complete, and Supermicro introduced a liquid cooled rack with 256 Vera CPUs. China orders · Modal · Supermicro

On the software side, NVIDIA released the full Run:ai platform as open source, launched Minerva, an AIOps platform for finding faults across AI factories, and introduced Confidential Bare Metal for running sensitive models on third party infrastructure. It also announced the OpenShell reference architecture for controlling AI agents, which BBVA plans to explore. Run:ai · Minerva · Confidential Bare Metal · OpenShell

European AI cloud, energy and cooling companies are adopting NVIDIA's DSX platform for AI factories, NVIDIA says. Armada will build modular AI factories on the DSX POD design and plans 500MW of distributed capacity with Johnson Controls in 2027, and DayOne says it will support the design in Southeast Asia. DSX in Europe · Armada · DayOne

China and export policy

DeepSeek plans to give Huawei computational data to help rework its Ascend AI chip design, and the two plan an open source model and software stack for Huawei hardware, The Information reports. DeepSeek also open sourced DeepInfer, its production inference stack, which supports Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. Cambricon's third quarter revenue of 3.53 billion yuan fell well short of forecasts as shortages of high bandwidth memory and limited access to advanced manufacturing held back production. DeepSeek and Huawei · DeepInfer · Cambricon

OpenAI has stopped pressing lawmakers for new chip export limits on China, Politico reports, which helps Nvidia as it fights proposals to attach chip limits to the annual defense bill. The US State Department sanctioned Shanghai Zhisuan Technology, MiraclePlus and three related Chinese entities after a former Google engineer's convictions for stealing code linked to Google's AI computing infrastructure. OpenAI · Sanctions

Chipmakers and suppliers

Lam Research reported record quarterly revenue of $8.22 billion, up 22% from the previous quarter, and forecast $8.6 billion for the December quarter. Texas Instruments' third quarter revenue rose 28% to $6.06 billion. BE Semiconductor Industries' third quarter orders rose 96% from a year earlier to €342.6 million, and its hybrid bonding orders have more than doubled compared with all of 2025. Lam Research · Texas Instruments · Besi

Qualcomm plans performance based job cuts worldwide as it moves resources toward AI infrastructure and data center chips; it did not say how many jobs. AMD and Intel will combine their PyTorch x86 CPU optimizations in the open source oneDNN library. Electronic system design revenue rose 12.2% to $5.71 billion in the second quarter, the ESD Alliance reports. Qualcomm · AMD and Intel · ESD Alliance

UMC and AP Memory are working on silicon capacitors for power delivery in AI systems, PUFsecurity's hardware root of trust received OCP S.A.F.E. approval for Caliptra based security, and Goodix launched fingerprint and secure element designs for AI edge devices. UMC and AP Memory · PUFsecurity · Goodix

Chip startups

Positron AI says General Compute will buy its Atlas inference systems and has committed to its next Titan platform, with Atlas offered to customers from early 2027. Dutch startup Photon Bridge raised a €10 million seed round for laser links between AI chips. Positron AI · Photon Bridge

Compute deals and neoclouds

Nscale and Anthropic are reportedly discussing a second compute deal of $45 billion, the same size as their first; terms could change. Nebius secured a $3.5 billion, 364 day facility from Citi for data centers and GPUs, published live GPU spot prices in Europe, and agreed a preliminary framework with HUMAIN and DataVolt for up to 50 MW at a planned campus in NEOM's Oxagon. Nscale and Anthropic · Citi facility · Spot prices · Oxagon

HUMAIN and NVIDIA plan up to 84 megawatts of AI infrastructure in Saudi Arabia, about 27,000 GPUs, subject to final agreements. G42 and Analog signed nonbinding memorandums with Yann LeCun's AMI Labs on world models and AI infrastructure. Perplexity will run its tools on customers' own infrastructure through Valarian's ACRA platform. HUMAIN · AMI Labs · Valarian

Korean neocloud Elice fell as much as 33% in its Kosdaq debut after a $151 million IPO, and Singapore's DayOne is weighing delaying its US IPO or cutting it to about $2.5 billion to $3 billion. CBRE agreed to buy London based Onnec for £460 million plus up to £300 million in earnouts. Elice · DayOne IPO · CBRE

Sea plans to sell its AI ready infrastructure as Sea Cloud. Shadeform launched Spot for reselling idle GPUs at a discount, and Fuse Energy opened UK reservations for a GPU service with the first 128 GPUs due in December. Duos Technologies signed a 10 year, 20 MW colocation deal with a US cloud provider worth more than $400 million. Sea Cloud · Shadeform · Fuse · Duos

New data center capacity

SK Telecom's SK Hyper is in talks to fund the first 5 gigawatts of a planned 15 gigawatts of AI data center capacity in South Korea by 2035, a first phase that would cost $350 billion. Public filings link General Catalyst and Aethel Energy to a proposed $66 billion AI data center in Wyoming, still at an early stage. IREN set up IREN Europe and secured a license for a planned 300MW campus in Badajoz, Spain. SoftBank founder Masayoshi Son plans to invest $50 billion in AI and data centres in Malaysia, Prime Minister Anwar Ibrahim said; SoftBank did not immediately comment. SK Hyper · Wyoming · IREN Europe · SoftBank

In Asia, Aolani signed nonbinding letters of intent for up to 500MW in Malaysia, STT GDC India will host up to 9,000 Neysa GPUs at Noida, SC Capital and NeoAI Cloud invested in a 40 MW site in Pohang, and Kazakhtelecom plans more than 2,000 additional GPUs. Aolani · Neysa · Pohang · Kazakhtelecom

In Europe, Greece inaugurated the €58.9 million Daedalus supercomputer, CommonAI chose Kao Data's Harlow campus for its ARIA funded Scaling Inference Lab, Polarise is building a 5 MW AI data center in Oslo with World Wide Technology, and Generative Bionics plans three physical AI data sites with AMD. Soluna and Applied Digital will develop a 50 MW wind powered site in Texas. Daedalus · CommonAI · Polarise · Generative Bionics · Soluna

Power

President Donald Trump said new large data centers will have to generate their own electricity and sell excess power to nearby communities or the grid. Venture Global signed agreements with GE Vernova and Baker Hughes for power units that could support up to 10 GW of compute off the grid. AEP and X-energy are studying a 960 megawatt nuclear project in Indiana with funding from Amazon, which plans to buy the power. Trump · Venture Global · Indiana nuclear

Google will fund about $600 million of Entergy Arkansas grid work for its West Memphis campus, and Voltus will give Google up to 1,070 megawatts from a virtual power plant on the PJM grid. Peak Energy raised $108 million for sodium ion storage for data centers, Vantage secured more than 1 GW of Heron solid state transformers, and Endeavour and Trevi Systems will use data center waste heat for desalination. Entergy · Voltus · Peak Energy · Vantage · Endeavour

NEXTDC halted trading and then cut its 2027 forecasts after a power delay at its Melbourne M3 site deferred about A$140 million in contracted revenue; its shares fell more than 40% and it paused new projects. NTT DOCOMO and NTT began a trial using AI to shift base station workloads with solar and battery power, and envelio's GPU power flow solver is in use at Czech grid operator EG.D. Trading halt · NEXTDC forecasts · NTT DOCOMO · envelio

Local opposition and rules

The Memphis City Council rejected an 18 month data center moratorium in a 6 to 6 vote after concerns including the expansion of xAI's Colossus. New York City Mayor Zohran Mamdani asked Governor Kathy Hochul to sign a bill pausing data centers that use more than 20 megawatts. An ABC News, Washington Post and Ipsos poll found 66% of Americans oppose new US data centers and 72% oppose one in their own community. Memphis · New York · Poll

Wisconsin's natural resources department will review the environmental effects of planned AI data centers after settling a lawsuit over Vantage's Port Washington campus, and the state's economic development agency signed at least five nondisclosure agreements with data center developers. A FracTracker and Public Citizen analysis found 35 Texas permitting actions after Governor Greg Abbott ordered a pause. Wisconsin review · Wisconsin NDAs · Texas

Google cleared an estimated 250,000 trees in Finland for a data center before environmental reviews were complete, the Guardian reports. Residents in Nova Scotia want more consultation on EverWind's planned campus for Anthropic. Amazon dropped plans for four data centers at George Washington University's Ashburn campus, and a client is considering turning San Francisco's empty SF Centre mall into a data center and AI lab. Finland · Nova Scotia · Ashburn · SF Centre

Senator Ron Wyden released a bill to tax data center operators and remove several tax breaks. The National Taxpayers Union says many states value data center servers well above their market value, ANSI called for coordinated data center standards, and a Vanderbilt Policy Accelerator report proposes using the Defense Production Act to protect non AI buyers of memory chips. Wyden · Server taxes · ANSI · Vanderbilt

Also

Overhaul says thefts of AI chips and data center equipment rose 34% since 2024, with more than $150 million stolen in 2026. JLL says AI data center construction is raising warehouse demand and rents near Dallas Fort Worth and Washington. Saxony and Arizona agreed to deepen semiconductor ties, and Park Place Technologies added GPU and liquid cooling monitoring for AMD, Intel and NVIDIA hardware. Overhaul · JLL · Saxony and Arizona · Park Place