Anthropic's inference revenue ($50M/megawatt vs $10-15M cost) creates a self-reinforcing loop: profit from serving models funds more training, which improves models, which raises…

OpenAI is on track to turn profitable in Q3 and has already contracted 40-50% of next year's incremental global compute, driving extreme centralization of compute resources at the…

SpaceX is building large-scale compute infrastructure and leasing it to Anthropic and OpenAI at premium prices ($25-50M/megawatt), recouping CapEx in a year, positioning itself as…

Google is a key infrastructure provider (TPUv7) and hyperscaler funding AI buildout through debt rather than cash flow, but captures less value than memory or model layers.

Specialized GPU/TPU cloud providers are emerging as the deployment layer for labs that want custom hardware without building datacenters themselves.

Frontier labs' revenue per megawatt has surged from negative to $50-100M+, enabling them to outbid all other compute consumers; they will absorb 70-80% of incremental compute by 2…

EUV lithography tools and their Zeiss mirrors are the hardest constraint in the semiconductor supply chain; expanding output takes years even with unlimited capital, creating mass…

TSMC's expansion is constrained by Carl Zeiss mirror production for ASML EUV tools, with ~100 tools/year by 2030 representing a hard physical limit that even massive capital infus…

Sophisticated trading firms are the highest-margin consumers of frontier model inference, generating far more value per token than model providers capture, but labs may internaliz…

Memory suppliers (Micron, SK Hynix, Samsung) are increasing prices much faster than TSMC or other supply-chain layers, capturing more value per unit of AI compute as HBM becomes c…

TSMC's value capture in the AI stack is surprisingly low compared to HBM suppliers, even though advanced packaging and leading-edge wafers are critical constraints.

SpaceX is building gigawatts of compute without a committed customer, then auctioning capacity to the highest-bidding frontier labs, recouping CapEx in one year at premium prices…

Meta and SpaceX are the only credible third compute providers; Meta uses its balance sheet and credit access to build speculatively, then chooses between internal monetization and…

Nvidia GPUs (Hopper, GB300, Vera Rubin) are the primary compute substrate for frontier labs; Jensen Huang has immense pricing power but has not yet fully exercised it despite labs…

As the dominant HBM supplier, SK Hynix sits at the tightest point of the AI hardware supply chain and is extracting maximum pricing power before capacity expands.

HBM and DRAM suppliers are the primary value capture point in the semiconductor stack; they can raise prices rapidly because they are the immediate bottleneck for GPU/accelerator…

Samsung supplies HBM to Nvidia/AMD and is a key source of advanced memory for China via smuggling/third-party channels, making it a geopolitical swing supplier.

China's semiconductor self-sufficiency push will inflect in 2027-28 as SMIC fabs and CXMT memory scale to millions of units/year, potentially adding 5-10 GW domestically in 2028 a…

China's leading memory manufacturer is ramping capacity in lockstep with SMIC logic fabs to enable a fully domestic AI hardware supply chain by 2028.

Chinese domestic fabs (SMIC, CXMT) will ramp to millions of units/year by 2028, adding 5-10 GW of domestically produced AI compute, though chip quality will lag US counterparts si…

Google is a key infrastructure supplier to Anthropic via TPUv7 and a major CapEx spender funding build-out through debt issuance and redirected buyback cash, but faces same crowdi…

Meta is building massive compute capacity on its balance sheet without committed customers, giving it optionality to sell to frontier labs at $25-50M/megawatt or use internally, m…

Proprietary trading firms like Jane Street extract far more value per compute unit than model providers by applying frontier models to high-frequency trading strategies, creating…