Why are Nvidia A100 GPUs that are already six years old reportedly contracted through 2029?
That fact changes the shape of the AI infrastructure debate. Commitments now extend into 2028, 2029, and 2030 across compute, memory, advanced packaging, land, and power.
But duration is not value. Scarcity can produce contracts and extraordinary margins, but it can also attract capital, provoke customer resistance, and require enormous spending before revenue arrives. The investment question is shifting from whether AI demand exists to which companies can convert physical bottlenecks into durable cash flow and attractive returns on capital.
The Bottleneck Has Moved Beyond the GPU
Jensen Huang identified “land, power and shell” as the next constraint as Nvidia, OpenAI, and SB Energy plan an 8 GW compute campus in Ohio. The first 4.25 GW is expected to use Nvidia’s DSX platform and begin coming online in 2028. The supporting energy buildout is expected to add at least 10 GW of new power.
Nvidia is using its balance sheet to help secure the sites and power needed for future accelerator sales. Capital allocated today is becoming tied to infrastructure that may not operate until 2028.
IREN’s CEO captured the physical mismatch: “Demand grows at the speed of software, but supply cannot.” Power connections, transmission lines, concrete, steel, and the workers required to build capacity cannot be provisioned like software.
That is persuasive as an industry argument. It is not sufficient as a company thesis.
A skeptical response to IREN made the distinction clear: “He says that a lot but still cannot get another deal.” Scarce connected power may be valuable, but scarcity does not prove that its owner can secure customers on attractive terms.
One promotional case claimed that IREN has secured NVIDIA and Perplexity, that Horizon 1 began generating approximately $120 million per quarter this month, and that annual recurring revenue could exceed $4 billion by year-end. Yet it provided no contract economics, financing assumptions, margins, or capital expenditure requirements.
Contracted volume matters. What the company spends to deliver it matters just as much.
Longer Commitments Point to a Longer Cycle
SanDisk reportedly has long-term agreements covering two-thirds of its 2028 output and minimum contracted revenue of $93 billion.
CoreWeave reportedly has contracts through 2029 for six-year-old Nvidia A100 GPUs. That is evidence against assuming useful GPU life must collapse rapidly as new generations arrive.
Applied Materials raised expected 2026 advanced packaging revenue growth from more than 50% to more than 70%. Its customer discussions extend through 2030.
TSMC warned that the industry is likely to face memory shortages and tight ABF substrate supply over the next few years. The marginal constraint can migrate through the system faster than new capacity can be added, moving among memory, packaging, substrates, networking, power, and land.
That suggests a longer infrastructure cycle. It does not mean every company attached to it will retain pricing power after capacity catches up.
Extraordinary Memory Margins Are Also a Warning
UBS estimates that conventional DRAM gross margins, including at Micron, could reach an unprecedented 95% by 2027 and surpass HBM margins. SanDisk reportedly projected approximately 80% adjusted gross margins through 2030, approximately 75% operating margins, and approximately 50% adjusted free cash flow margins.
Long-term agreements can improve visibility and make this cycle last longer than investors using an old spot-price template might expect.
But projected gross margins of 80% or 95% create their own counterforce. Economics that attractive can draw new capital and intensify customer resistance. The facts support the possibility of a longer memory cycle, not the conclusion that memory has stopped being cyclical.
Scarcity is often most persuasive near the point when the market begins treating it as permanent.
Adoption Remains Concentrated
Ramp reported that in July, the top 1% of US businesses spent a record median of $7,400 per employee per month on AI. That compares with $650 for the top 10% and $11.95 for the median business.
At the same time, only 7% of companies reportedly have fully implemented AI. About 30% reported increased labor productivity, while 56% had an AI account.
These figures do not negate infrastructure demand. They identify the unresolved economic test. The industry is making long-duration commitments before AI has been fully implemented across most companies. Ultimately, productivity must justify the capital committed across accelerators, memory, packaging, networking, land, and power.
The evidence supports a longer infrastructure cycle, not indiscriminate ownership of everything labeled AI. Returns depend on who can turn physical scarcity into contracted cash flow without destroying returns through capital intensity.