Sold Out
From foundry quotas to rationed compute, AI’s supply chain has run out of room.
TSMC’s second-quarter earnings earlier this month capped several weeks of signals from across the AI supply chain. Samsung’s foundry division imposed quotas on its advanced process nodes. Google informed Meta that it could not supply the Gemini capacity Meta had sought to purchase. AMD and Rackspace formalized a governed compute partnership for enterprises facing infrastructure bottlenecks. Each of these developments sits at a different layer of the stack. Together they describe a single structural condition in which demand for AI infrastructure exceeds supply from the fabrication line to the finished compute product.
Fully booked
TSMC reported record revenue of $40.2 billion for the second quarter of 2026, a 36 percent increase from the same period a year earlier. Net profit rose 77 percent to $22.4 billion. High-performance computing, the segment built on AI accelerators for cloud data centers, accounted for 66 percent of wafer revenue and grew 20 percent from the previous quarter alone.
TSMC has sold out its CoWoS advanced packaging capacity through year-end, with lead times extending into 2027. CoWoS bonds a processor to its high-bandwidth memory on a single interposer, and every high-end AI accelerator that Nvidia and AMD ship to data centers must pass through the process. The foundry’s 3-nanometer production node carries a similar constraint, with allocation fully committed through 2026. The company raised its full-year revenue growth forecast above 40 percent.
This constraint spans both major foundries. Samsung’s foundry division implemented a quota system on select process nodes, prioritizing existing customers as orders exceeded capacity. Orders have fully booked the division’s 4-nanometer lines through next year, with some 8-nanometer production running near full utilization. Demand has shifted from smartphone application processors to AI accelerators and high-performance computing chips. Samsung now fabricates autonomous driving silicon for Tesla and inference chips for Groq, and has entered early discussions with Anthropic about a custom chip on its 2-nanometer process.
A $1B waiting list
Google informed Meta in March that it could not supply the full Gemini capacity Meta had sought to purchase. This shortfall disrupted several of Meta’s internal AI projects and prompted the company to instruct employees to conserve AI tokens. Meta had relied on Gemini for content moderation and scam detection, tasks for which Google’s models proved more capable than Meta’s own Llama family.
Google’s own capacity position explains that rationing. The company pays SpaceX $920 million per month for access to 110,000 Nvidia GPUs as bridge capacity for Gemini Enterprise. Google Cloud’s quarterly revenue topped $20 billion in the first quarter, up 63 percent from a year earlier, and its backlog of signed but undelivered contracts roughly doubled to approximately $460 billion. Google also tightened Gemini usage limits for subscribers, switching from prompt-based caps to a system weighted by request complexity and conversation length. Alphabet has committed $180 to $190 billion in capital expenditure for 2026.
Meta has accelerated its own infrastructure in response. The company shifted content moderation and safety workloads to Muse Spark, a proprietary model from its Superintelligence Labs division, to reduce dependence on a competitor that doubles as its compute supplier. Meta’s 2026 capital expenditure guidance reached $115 to $135 billion, and the company cut 8,000 positions while reassigning 7,000 employees to AI-focused roles.
Alternative supply
AMD and Rackspace signed a definitive agreement in June for the phased deployment of 30 megawatts of AMD-powered compute across Rackspace’s global data centers, beginning in late 2026 and running through 2028. The deployment incorporates AMD Instinct GPUs, including the MI355X and MI350P, alongside EPYC CPUs, and targets regulated enterprise workloads with healthcare providers among the first organizations to express interest. Rackspace frames the offering as a governed stack with a single operator accountable for every layer, from accelerated compute to AI inference in production.
AMD’s parallel partnership with Nutanix extends the approach, aiming to build an open, full-stack AI platform for agentic workloads that integrates ROCm software and AMD hardware into Nutanix’s infrastructure. AMD cites survey data showing that 64 percent of organizations identify data and infrastructure bottlenecks as the primary obstacle to AI deployment, a figure that the TSMC and Google capacity data corroborate from the supply side.
Infrastructural leverage
TSMC’s gross margin reached 68 percent in the quarter. Sold-out capacity confers pricing power, and the companies that control foundry capacity, advanced packaging, and power infrastructure now hold more structural leverage over AI’s trajectory than the companies training the models. AMD and Samsung have begun absorbing the demand overflow, while Meta and Google spend hundreds of billions to build the capacity neither can buy from the other. New fabrication capacity takes years to build, and new packaging lines take nearly as long to qualify. The current investment cycle measures the distance between demand and the pace at which supply can respond.