Analysis
Four separate disclosures inside a single week describe the same underlying story from four different angles: OpenAI revealed a self-built, $30 billion-plus data center campus in Georgia on July 22; Nvidia prepaid Amkor $1.5 billion on July 24 to build U.S. advanced packaging capacity; Samsung and SK Hynix moved the same day to sign large HBM memory megadeals with U.S. tech firms, building on an already-disclosed roughly $880 billion joint expansion plan; and Intel separately raised its own 2026 capex guidance from $18 billion to more than $20 billion, only to see its stock fall 11% on investor skepticism about exactly that spending.
Read individually, each is a company-specific capital allocation decision. Read together, they describe an AI supply chain where compute, power, packaging and memory are all being locked down simultaneously by different players, rather than one bottleneck getting solved before the next emerges. That's a meaningfully different pattern than the 2023-2024 buildout, when GPU scarcity was the singular constraint everyone pointed to; by mid-2026, power delivery, advanced packaging and HBM memory have each become independent choke points commanding their own multibillion-dollar prepayment and megadeal structures.
The competitive dynamic cuts in an unusual direction: it's not just the AI labs competing with each other, it's every layer of the supply chain gaining leverage at once. Samsung and SK Hynix's undisclosed but described-as-"very large" HBM deals suggest memory pricing power has shifted toward suppliers after years of commoditized DRAM cycles. Amkor's roughly 10% stock pop on the Nvidia news shows packaging suppliers now command the kind of investor attention historically reserved for the GPU makers themselves. And OpenAI choosing to self-build Project Camellia rather than lease more cloud capacity signals even the best-capitalized labs see owning infrastructure, not renting it, as the more durable long-term position.
“Amkor's roughly 10% stock pop on the Nvidia news shows packaging suppliers now command the kind of investor attention historically reserved for the GPU makers themselves.”
The public-markets reaction to Intel's capex raise is the tell that this compounding spend is starting to worry generalist investors, not just infrastructure specialists. A chipmaker beating earnings by $1.7 billion and still dropping 11% because of a $2 billion capex guidance increase shows the market has become genuinely capex-fatigued -- rewarding revenue growth less than it's punishing open-ended spending commitments, even from companies with real customer traction behind the numbers.
For venture investors, the pattern argues for underwriting AI infrastructure bets on capital efficiency and defensible unit economics rather than pure growth velocity, since public markets are now visibly doing exactly that to the biggest, best-capitalized players in the space. For founders, it's a reminder that every layer of this stack -- compute, power, packaging, memory -- is being fought over as a distinct, investable category in its own right, not a commodity input.
Watch whether Q3 earnings from Google, Amazon and Meta show similarly rising capex guidance meeting similarly skeptical stock reactions, whether Samsung and SK Hynix disclose actual dollar figures behind this week's HBM megadeals, and whether any of these four moves shows measurable signs of easing the specific bottleneck it targets within the next two quarters.
