Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on Friday at WAIC in Shanghai, and the semiconductor sector spent the session repricing the hyperscaler capex thesis around it: the Philadelphia Semiconductor Index closed the week down roughly 20% from its late-June record, a bear market that erased about $3.3T in combined market value, with equipment makers like Applied Materials among the hardest hit. The shock replays January 2025's DeepSeek moment — a materially cheaper, competitive frontier model undercuts the case for GPU-training capex at current scale — though SemiAnalysis argued the same day that Kimi K3's scale actually requires more hardware, not less: the model doesn't fit on a single Nvidia DGX B200 even at FP4 precision, and needs a minimum 64-chip scale-up domain to serve. Watch whether Kimi K3's real-world serving costs over the coming weeks vindicate the panic-sell or the Jevons Paradox read.
Private Companies
xAI
launch
Elon Musk said on X July 18 that xAI's next flagship model, Grok 4.6, is a 2-trillion-parameter system finishing its initial training run, claiming it could exceed Kimi K3's benchmarks while matching the serving speed of the current 1.5-trillion-parameter Grok. The claim landed a day after Moonshot reset the frontier-model leaderboard. Grok 4.6 hasn't shipped and the efficiency claim is unverified, but keeping pace with Moonshot's scale — rather than ceding the largest-model title to a Chinese lab — is now a competitive question for xAI as much as a technical one.
SambaNova
other
Upper90 issued a $400M loan to AI cloud startup General Compute collateralized by SambaNova SN50 inference chips, TechCrunch reported July 17 — the first financing deal built on inference-specific silicon rather than Nvidia training GPUs. It extends the GPU-backed lending pattern Upper90 has used elsewhere to a non-Nvidia inference vendor. For SambaNova, having its chips treated as bankable collateral is a maturity signal that argues its hardware carries resale value independent of the company's own balance sheet — a bar few non-Nvidia inference chipmakers have cleared.
Public Markets
Google is reportedly developing a "Frozen V2" chip that hardwires Gemini's model weights directly into silicon, targeting 6-10x efficiency gains by 2028 — a new axis of hardware specialization beyond general-purpose TPUs.
Emerging
Hut 8 signed a second 15-year, $9.8B lease with an existing investment-grade tenant, Reuters reported July 20, fully commercializing its 1-gigawatt Beacon Point campus in Texas at 352 MW of IT capacity and doubling that customer's contracted footprint there to 704 MW. It brings Hut 8's total contracted AI data-center capacity to 949 MW, backed by 1,330 MW of utility capacity, with aggregate base-term contract value across its portfolio reaching $26.6B. A former bitcoin miner converting power and site assets into AI data-center capacity at this scale is now a repeatable playbook, not a one-off — watch which other crypto-to-AI conversions land similarly sized anchor tenants.
Researchers built the largest probabilistic computer yet, networking 18 FPGAs into a single machine with 1 million p-bits, up from 7,200 p-bits in 2023, according to an arXiv paper. The system runs Gibbs sampling at over a trillion flips per second, breaking probabilistic computing's single-chip ceiling by distributing an Ising machine across networked commodity hardware. P-bit accelerators target stochastic optimization problems without quantum computing's cooling and error-correction overhead — watch whether probabilistic silicon reaches commercial deployment before neuromorphic or quantum rivals do.
Watch This Week
Wednesday, July 22
SK Hynix reports Q2 2026 earnings, the first read on memory pricing power since CXMT's IPO and this week's sector selloff (its investor call follows July 29). Texas Instruments also reports Q2 earnings the same day, a bellwether for broad-based industrial and auto chip demand.
Thursday, July 23
Intel reports Q2 2026 earnings, watched for whether 18A/foundry progress is translating into external customer wins. STMicroelectronics also reports Q2 earnings, a read on auto/industrial demand and SiC power capacity.