Leopold Aschenbrenner's AI-focused hedge fund, Situational Awareness LP, was forced to liquidate nearly all of its public equities book this week after leveraged bets on AI-infrastructure names — SK Hynix, Sandisk, Bloom Energy, Nebius among them — fell more than 30% over the past month on investor concern that hyperscaler capex isn't converting into revenue fast enough. The fund, up 439% through June and peaked near $45B in assets, sold the bulk of its public holdings in a block trade to Ken Griffin's Citadel and now runs roughly $10B, mostly private positions including a reported $5B stake in Anthropic. It's the first forced-deleveraging event inside the AI-infrastructure trade itself — proof the same capex-cycle uncertainty producing today's Microsoft and Lam Research rallies can just as easily unwind a book built on the other side of that bet. (WSJ)
Private Companies
ModularACQUISITION
Qualcomm closed its $3.9B all-stock acquisition of Modular on July 29, issuing up to 19.2M shares to the startup's equityholders. Founder Chris Lattner — creator of LLVM, Clang, and Swift — becomes EVP of Advanced AI Software and Platforms at Qualcomm, while Mojo, MAX, and Modular Cloud continue as products letting AI workloads run unmodified across CPUs, GPUs, NPUs, and custom silicon. It's a bet that the compiler and runtime layer, not raw silicon, is where a challenger finally breaks Nvidia's CUDA lock-in. (PR Newswire)
WEKAPartnership
Andromeda, a platform routing AI workloads across 50+ compute providers, is standardizing on WEKA's NeuralMesh — including the GPU-native NeuralMesh Axon — as the storage layer under its managed GPU clusters. Andromeda cites sustained throughput above 400GB/s and 6M IOPS per cluster, environment startup times cut from three minutes to 30 seconds, and a first Axon deployment stood up in 10 minutes from bare metal. It's another neocloud treating NeuralMesh as default infrastructure regardless of which hyperscaler sits underneath — vendor-agnostic positioning that matters as buyers multi-source capacity. (AIwire)
Public Markets
A fiscal Q4 beat and a Q3 guide well above consensus ($8.1B revenue vs. $7.13B expected) drove Lam's best day since 1999 and lifted the whole semicap complex — Applied Materials, AMD, and Micron all rose double digits — signaling fab equipment demand is still inflecting upward. (CNBC)
Emerging
A silicon "governor" for export-controlled AI chips. Princeton researchers propose Hardware Mechanisms to Dynamically Throttle AI Performance — microarchitectural circuits built into a GPU's memory subsystem (cache way-masking, credit-based rate limiting, bank arbitration) that cut measured AI performance by up to 80% for under 10,000 flip-flops of added silicon. It's a hardware-native alternative to export licensing and software guardrails: a governor chipmakers could embed to make usable AI throughput provably and cheaply capped, relevant to how Nvidia, AMD, and hyperscaler silicon teams might design tunable performance limits into future export-controlled accelerators. (arXiv)
China's compute build-out keeps scaling regardless of export enforcement. Bloomberg reports DeepSeek is planning a roughly 1-gigawatt data center in Ulanqab, Inner Mongolia — built partly in-house, partly leased — targeting partial capacity by late 2027; chip sourcing isn't confirmed, though DeepSeek has previously been linked to export-controlled Nvidia Blackwell hardware. Ulanqab's cheap power and cold climate are turning it into a hub for China's parallel build-out, with Z.ai reportedly completing a domestic-chip 1GW cluster there this month. (Bloomberg)
Watch This Week
Mon Aug 3
Palantir reports Q2 2026 results after market close; consensus expects ~81% YoY revenue growth, a read on enterprise/government AI-infrastructure demand pull-through ahead of the week's chip prints. (StockTitan)
Mon Aug 3
onsemi reports Q2 2026 results after market close, a power/analog bellwether increasingly tied to AI-datacenter power delivery. (StockTitan)
Tue Aug 4
Tue Aug 4
Astera Labs reports Q2 2026 results after market close, the same day as AMD — a direct read on PCIe/interconnect retimer attach rates inside the same GPU-cluster buildout. (StockTitan)