The Chip Boom Has Moved to the Chokepoints

The most revealing number in the semiconductor boom right now is not a capex figure or a stock price. It is 192 — the gigabytes of HBM memory that Nvidia's next-generation Blackwell accelerator requires, up from 141 in the H200 and roughly 80 in the H100 ^[1]. That geometric climb, not any forecast of server-count growth, is what has quietly redirected hundreds of billions of dollars. The money is no longer chasing generic capacity. It is hunting the specific chokepoints that now gate how fast AI can be deployed — and the market has already begun to price the difference. In the first half of 2026, memory and hardware suppliers surged: Sandisk rose 858%, Micron 300% to a $1.1 trillion valuation, Intel 280% ^[2]. Over the same six months, infrastructure suppliers drove the Philadelphia Semiconductor Index up more than 90% ^[3]. Meanwhile, in June alone, Microsoft, Nvidia, and Apple — three of the market's most valuable companies — shed $2.3 trillion in market value ^[3]. Software companies collapsed: Intuit down 60%, Adobe and Workday down 37% each, despite growing revenue ^[2]. The market is drawing a line between the companies that own the bottlenecks and the companies that pay for passage through them. The mechanism that makes this more than a cyclical squeeze is architectural. Each generation of AI accelerator demands geometrically more high-bandwidth memory: H100 carried roughly 80GB, H200 jumped to 141GB, Blackwell reaches approximately 192GB ^[1]. Memory demand per chip compounds even if the number of servers shipped plateaus. Micron's growth thesis has shifted accordingly — from counting AI servers to tracking memory density per unit — and CEO Sanjay Mehrotra now calls the moment an inference inflection, describing memory as a strategic asset whose supply cannot be brought up easily ^[4][1].

This is inference inflection. — Sanjay Mehrotra

That logic — the bottleneck is structural, not temporary — now anchors the pitch documents at TSMC and SK Hynix, and sits at the center of Korea's national chip initiative ^[5][6]. And it is reshaping how capital flows into the industry, through mechanisms that did not exist at this scale two years ago. The most direct is the contract. Micron has abandoned the industry-standard one-year term and shifted a third of its NAND and 20% of its DRAM volume onto five-year take-or-pay agreements, locking in roughly $100 billion across 16 customers ^[7][8]. The memory shortage is no longer a spot-market spike a buyer can wait out; it is a multi-year lock-in. Apple has taken the same logic further: sitting on $123 billion in cash, it has pre-paid for multi-year HBM supplies from SK Hynix and Micron through 2026 and booked more than half of TSMC's initial 2-nanometer capacity ^[9]. It is not buying chips on the open market. It is buying the right to go first.

Our AI investments and full stack approach are lighting up every part of the business — Sundar Pichai

At the design level, the land-grab is even tighter. Nvidia's Jensen Huang traveled to Korea in June and sealed a multi-year partnership with SK Hynix to co-develop next-generation HBM specifically for the Vera Rubin AI accelerator ^[10]. Nvidia is no longer simply purchasing memory from a supplier — it is designing the bottleneck component alongside its largest producer, locking in the critical-path supply before competitors can reach it. Every major hyperscaler is now building its own custom AI silicon rather than relying on commodity GPUs. Broadcom controls roughly 95% of the co-design market, with contracts for Google's TPU through 2031, Meta through 2029, and a $30 billion deal with Apple ^[11][12]. Marvell architects Amazon's Trainium and Microsoft's Maia ^[12]. OpenAI and Broadcom unveiled Jalapeño — a custom inference chip designed in nine months that cuts inference costs roughly 50% against GPU clusters ^[13]. These ASICs deliver 40 to 60% total cost of ownership reductions versus GPU clusters, and Deloitte projects the custom AI chip market will exceed $50 billion in 2026 ^[14]. The capital is flowing toward bespoke accelerator silicon, not commodity compute. The national scale of the shift is visible in South Korea's $2 trillion decade-long AI and chip initiative, centered on an 81 trillion won packaging hub and four new fabs. SK Hynix raised $26.5 billion in a record Nasdaq debut explicitly to fund HBM and advanced packaging ^[6]. TSMC committed to a US-based advanced packaging facility and raised its 2026 capex to $54 billion, above the $47.8 billion consensus, with Chairman C.C. Wei confirming the packaging bottleneck is the target, not just more wafer capacity ^[5].

strong AI-related demand and growing computing requirements continue to support solid demand for chips manufactured with its leading-edge process technologies. — TSMC

The chokepoint hunt has even reached components that were invisible a year ago. Multi-layer ceramic capacitors — MLCCs, the tiny passive components that regulate voltage — have become the third most expensive item in an AI server after GPUs and memory, with some racks requiring 600,000 of them ^[15]. Lead times exceed 20 weeks and spot prices in Shenzhen's Huaqiangbei electronics market have risen tenfold ^[15]. Credo Technology, which holds 80% of the Active Electrical Cable market for AI data center interconnects, is being targeted by major funds as a structural AI tailwind play ^[16]. The bottleneck thesis has migrated from the headline components to the wiring and the capacitors. All of this is real. But all of it rests on a single unproven bet: that AI demand persists at anything like the levels that justify it. The counter-evidence is not marginal. Hyperscaler gross leverage has doubled from 0.9x to 1.8x in two quarters ^[17]. UBS has tracked $800 billion in US credit funding AI capital expenditure ^[17]. Alphabet raised up to $84.75 billion in the largest equity capital markets deal in history, with 2026 capex of $180 to $190 billion potentially rising to $300 billion in 2027 ^[18]. And 95% of organizations report no measurable return on their AI investments ^[17].

Labs like OpenAI and Anthropic are going to have to increase prices, they're going to have to cut costs, or there's going to be a big bubble explosion. — Yann LeCun

Enterprises hit by usage-based token pricing are already routing to cheaper open-source models, including DeepSeek, and building in-house GPU compute to escape the meter ^[19]. A class-action lawsuit alleges Samsung, SK Hynix, and Micron used the HBM shift as a pretext to wind down legacy production and artificially inflate DRAM prices 700% — suggesting the bottleneck narrative may partly mask coordinated supply restriction ^[20]. And AI data center expansion is hitting power-grid bottlenecks while open-weights Chinese models and edge hardware threaten to shift inference from the cloud to the device, undermining the centralized demand that justifies the capex ^[21]. The land-grab for chokepoints is not a mirage. The contracts are signed, the fabs are breaking ground, the five-year take-or-pays are binding. But the architecture of the bet is unmistakable: the industry has borrowed against a demand curve that has not yet proven itself durable. If AI demand holds, the chokepoint owners will have positioned themselves brilliantly. If it does not, the bottleneck becomes an overhang — and the same leverage that funded the land-grab will fund the unwind.