Meta negotiates $10B two-year compute lease with Anthropic, Alibaba's Qwen3.8 Max lands at 2.4T parameters w/DeepSeek V4 close behind, and PC makers lock up CXMT DRAM through 27 as Washington watches
Meta wants to be a cloud vendor. It’s in early talks to lease up to $10 billion of surplus compute to Anthropic over two years. Elsewhere, Alibaba previewed Qwen3.8 Max at 2.4 trillion parameters and DeepSeek V4 nears launch with dynamic pricing, extending China’s open-weight compression streak. Separately, PC vendors have locked CXMT DRAM orders through 2027 even as US lawmakers push to ban CXMT and YMTC purchases.
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
Presented by …(this spot is open: reply to this email to talk sponsorship)
Meta Turns Surplus Capacity Into Cloud Business With Anthropic Compute Lease
Meta and Anthropic are in early talks for a compute lease worth up to $10 billion over two years, The New York Times reports, with Meta monetizing excess data-center capacity as a merchant cloud entrant (Bloomberg Tech). The deal signals Anthropic systematically diversifying away from single-source compute dependence, a posture visible across multiple supplier relationships.
Meta’s move reframes its CapEx-heavy infrastructure build as a revenue-generating cloud business, not purely internal tooling.
If signed, the deal would be Meta’s first major external compute-leasing contract; watch Meta’s Q2 2026 earnings for any infrastructure-revenue line disclosure as the concrete next checkpoint.
Vik:Meta had been using almost all of its compute infra internally. Apparently, Zuck got such amazing offers for compute, he decided, why not rent it out and collect the cash. Anthropic is hungry for compute, so why not sell to them? Bird in hand is worth two in the bush. Interesting to see if other hyperscalers will follow.
Austin: Vik, wasn’t Meta the only hyperscaler not renting out GPUs? Microsoft, Google, and Amazon all sell cloud. Meta doesn’t. Until now. btw, the rent checks from GPU tenants are nice, but the discipline is nicer… every internal GPU project now has to beat a market price. No clear ROI? Rent the silicon to Anthropic instead.
Not that Meta lacks internal uses, from ad-model gains today to LLM features tomorrow. But discipline is good. Related: my conversation with Meta’s Matt Steiner, who runs monetization infrastructure, ranking, and AI foundations.
Qwen3.8 and DeepSeek V4 Extend China’s Open-Weight Pricing Squeeze
Following Kimi K3’s chip selloff, two more Chinese releases landed within days: Alibaba previewed Qwen3.8 Max at 2.4 trillion parameters, claiming it is second only to Anthropic’s Claude Fable 5, while DeepSeek V4 nears full launch with peak-valley dynamic pricing, together broadening the open-weight compression trade from a single event into a sustained pattern.
Alibaba shares rose 5.4% on the Qwen3.8 preview (Bloomberg Tech), though DigiTimes notes the claim carries no published benchmarks.
Moonshot suspended new Kimi K3 subscriptions within 48 hours of launch as GPU capacity hit limits (SCMP); the company is targeting an IPO within six months.
Three trillion-scale open-weight Chinese models now in market simultaneously; next hard checkpoint is DeepSeek V4’s full release, expected imminently per 36Kr.
Vik:Open weight models are landing every day now. This opens up the use of AI models outside of what compute the big dogs like OpenAI and Anthropic can get their hands on. These kinds of models are only good for AI CapEx; we need more compute. What is more interesting is: what kind of regulations and security measures are now needed on frontier models like K3, Qwen 3.8 Max, Fable and GPT 5.6 Sol? I read a very interesting essay by Dean Ball on this topic that has been making rounds on X.
Austin:Open weights models are great for enterprise inference demand; just need more competitive American open weight alternatives. Also, I’m tired of the OpenAI vs. Anthropic story. It’s Zoolander and Hansel in a walk-off… two beautiful rivals who only see each other. Let’s get more models and crowd the runway. (double entendre, see what i did there)
PC Vendors Book CXMT Capacity Through 2027 as Ban Calls Mount
Following the $8.6B IPO, PC makers are now fighting for CXMT DRAM allocation with orders reportedly locked through 2027, tightening commodity supply even as US lawmakers push to ban CXMT and YMTC chip purchases outright. (TechPowerUp)
CXMT capacity is tracking to near Micron’s output levels by end-2026, with a process node transition targeting DDR6 before commercialization. (TweakTown)
US legislators have formally urged a purchase ban on CXMT and YMTC chips. (bloomingbit)
Smaller OEMs are already being squeezed out of allocations; the concrete next checkpoint is whether the US ban proposal advances to committee markup before Q3 2026.
Vik:PC makers can’t get memory from the big 3 who are too busy allocating 3x wafers to HBM. Sure, DRAM sales are more profitable, but it is important for the big memory guys to satisfy their LTAs with their large customers. PC makers are left high and dry to go to China, but then they get blocked by regulatory bans. No winning either way for PC guys.
Quick Takes
Compute
Foxconn reportedly won a $52B SpaceX contract for 13,000 Nvidia GB300 server racks, though Elon Musk publicly branded the report ‘fake news.’ (DigiTimes)
Samsung SDS launched Korea’s first NPU-as-a-Service on FuriosaAI’s RNGD chip via Samsung Cloud, supporting sovereign-AI ambitions. (Chosunbiz)
Memory
Kioxia shares hit limit-down after a US jury ordered it to pay Viasat $229 million in a patent dispute, halving its market cap in a month. (nhk.or.jp, Seoul Economic Daily)
Foundry & Packaging
ASML lifted its 2026 revenue target to €45B and plans a 30% EUV capacity surge, while reportedly pushing Low-NA price hikes that frustrate TSMC. (EE Times, Tom’s Hardware, WinBuzzer)
TSMC accelerated its Arizona buildout to $265B, citing ‘strong multi-year’ AI demand and 4-5x higher US fab construction costs, per CFO. (CNBC, Bloomberg Tech, Reuters)
Rapidus partnered with Cadence to integrate agentic-AI design tools into its 2nm foundry process, targeting shorter development cycles. (Nikkei Asia, EE Times Asia, Electronics Weekly)
Presented by …(this spot is open: reply to this email to talk sponsorship)
Key Data
Funny
Felt like just yesterday when Hunterbrook found hidden HTML with Anthropic partnership details — which was supposedly not real. Now some AMD Senior Director has checked in Github code? LOL
