Meta planning to sell AI compute has triggered a market selloff in neoclouds, and in general, AI stocks. Oracle questions profitability. MOAR memory news. ++
Let’s get into it. — Austin & Vik
Quick hits, high signal. Takes from semi industry experts. Sign up for free daily updates!
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player
Meta Plans Cloud Business, Selling AI Compute and Models
Meta Platforms Inc. is developing plans for a cloud infrastructure business to sell access to AI computing power and models. This initiative aims to generate revenue from excess computing power, potentially offering access to AI models hosted on Meta’s infrastructure, similar to AWS Bedrock, or selling raw computing capacity like CoreWeave Inc. The plans are part of Meta Compute, an internal initiative led by Santosh Janardhan, Daniel Gross, and Dina Powell McCormick. (bloomberg.com)
Vik: This piece of news has led people into thinking that compute has peaked. In reality, it is a pathway for Meta to be more profitable going forward, in spite of all their capital investments. Meta has historically been a customer of CoreWeave, and with them selling compute, the market views this as eating neocloud lunch, causing their stocks to drop.
Micron Becomes Most-Traded US Stock with $70B in Daily Turnover
Micron recently became the most-traded stock in the US, reaching $70 billion in daily turnover. The trading volume reflects investor sentiment that memory chips are a critical and constrained component for AI compute infrastructure. This has led to a re-evaluation of memory from a cyclical commodity to a strategic asset with significant pricing power, particularly for domestically manufactured chips. (DigiTimes)
Vik: Yeah, we know that memory is important. More people buying/selling memory stock than ever. People are so skittish on memory right now. Everybody is frantically looking for the peak that even the slightest gust of wind, like the tweet below sets off a market sell-off.
For what it’s worth, I think Andrew is referring to Core Automation, an openAI spinoff that is rethinking how AI works.
Vik: Watch the video. It’s a nice homely robot folding clothes and doing laundry. If it can clean up after young kids, its worth the $8K. Anybody objecting to the price tag needs to understand how messy life with young kids can be, and how a well to do family would buy one. Also, these things get cheaper over time.
Micron Technology and General Motors have signed a Strategic Customer Agreement to secure long-term supply of memory and storage semiconductors for GM vehicle platforms. The deal covers high-performance DRAM and NAND components used in next-generation GM vehicles, giving GM committed supply access while Micron locks in an automotive customer relationship. (Micron, Reuters)
Vik: Geez Louise! Auto industry is locking up memory supply now too?!
Nvidia, Partners Pledge US Manufacturing Investment
Nvidia and an unspecified group of domestic partners announced commitments to build semiconductor and AI infrastructure in the United States, targeting manufacturing capacity, supply chains, energy infrastructure, and workforce development. The company framed the initiative around applications in healthcare, scientific research, and industrial productivity, though no dollar figures, partner names, or production timelines were disclosed in the announcement. (Nvidia News)
Vik: Here is Nvidia’s “made in America” map. Montana, South Dakota, Louisiana, West Virginia, and Vermont have no hope. No NVIDIA green for you!
Oracle Questions Returns on AI Data Center Spending
Oracle warned that its investment in AI data centers may not pay off, raising concerns about the economics of large-scale AI infrastructure buildout. The caution comes as Oracle is a principal partner in Stargate, a joint venture with OpenAI and SoftBank to build AI data centers across the U.S., including a facility in Abilene, Texas. Stargate has received promotional backing from President Trump and represents one of the largest committed AI infrastructure programs in the country. (Bloomberg Tech)
Vik: Oracle has a $300B Stargate contract with OpenAI and the Oracle’s company filing lists nonpayment as one of the risks. Stock has been dropping pretty steadily this month too.
Oxmiq Raises $35M Series A, Expands to Data Center Design
Oxmiq Labs, the GPU IP startup founded by former Intel and AMD executive Raja Koduri, has raised $35M in a Series A round, bringing total funding to $60M. The company, which develops GPU hardware and software intellectual property, will use the capital to expand beyond GPU IP into broader data center design. (EE Times)
Vik: Oxmiq basically builds silicon IP and software for the entire AI stack that you can just license and build. Raja Koduri is a highly respected individual and an industry veteran who also posts great insights on X.
ASE Reportedly Raises Advanced Packaging Prices by Over 20%
ASE is reportedly raising its quotes for advanced packaging by more than 20 percent, driven by high demand for AI components. The price increase highlights acute supply constraints and the critical role of packaging in AI compute scaling. (TrendForce)
VikL Prices only go one way these days. Up.
TSMC Q2 Gross Margins Near 70% on AI Chip Demand
TrendForce reports that TSMC’s gross margins for the second quarter are approaching 70 percent, driven by strong demand for its 3nm and 5nm nodes used in AI accelerators. The company’s third-quarter revenue is projected to grow over 10 percent sequentially, reflecting its pricing power amid tight CoWoS packaging supply. (DigiTimes)
China’s DeepSeek Raises $7.4B to Compete with Frontier AI Models
Chinese AI firm DeepSeek has raised $7.4 billion to develop models competitive with Anthropic’s Mythos, while Zhipu’s GLM-5.2 has shown strong performance in cybersecurity benchmarks. The developments signal a narrowing capability gap between US and Chinese frontier AI models, as ByteDance also pursues in-house CPU designs for 2027. (South China Morning Post)
Important Tweet
Amazon is shifting away from externally sourcing silicon for its products, and instead plans to adopt a customer owned tooling (CoT) approach with Alchip to develop its own chips for kindle, fire TV, Alexa, etc. Expected to begin in 2027. Annual shipments expected around 40M units.
