The Wall Street Journal has taken aim at one of the pillars of the AI hardware rally: the massive long-term supply deals now being signed across the memory industry.
Its argument is that agreements for silicon struck during scarcity have a long record of failing to hold, with Microchip Technology's collapsed pandemic-era commitments the cautionary tale.
On the history, there is little to dispute.
But the comparison glosses over how the current crop of strategic customer agreements (SCAs) differs from the contracts that unravelled in past cycles, and memory vendors are acutely aware of that history.
Contracting less than they could
The first difference is restraint.
Micron has signed 16 SCAs covering roughly 20% of its DRAM output and about a third of its NAND volumes, backed by $22 billion in customer commitments.
Given high-bandwidth memory capacity is sold out through 2026 across all three major suppliers, vendors could plausibly contract close to 100% of output.
They are choosing not to, and the contracted volumes appear to be based on current rather than future output.
That matters because the classic failure mode, over-shipping into customers until everyone gets hurt, only materialises when contracted amounts exceed real demand.
If commitments sit well below underlying demand, simply meeting the contracts does not create over-shipment.
Penalties that do not fade
The second difference is contract structure.
Historically, penalties for walking away from long-term silicon agreements have diminished over the life of the contract, making a late-cycle break relatively cheap.
The new SCAs appear to keep penalties substantial right through the contract tails.
That is precisely when customers would be most tempted to break terms, since memory pricing shows minimal likelihood of declining into 2028, while the new fabs now under construction, including Samsung's P5 plant and SK Hynix's M15X, necessarily add supply risk toward the end of the decade.
Static penalties make defection at that point far less attractive.
No Covid-style hoarding
The third difference is the absence of inventory building.
The pandemic saw a one-off surge in demand for remote-working infrastructure, compounded by corporations over-ordering and stockpiling chips as insurance against supply chain shortages.
When that demand normalised, the contracts signed at the peak became liabilities.
There is no equivalent hoarding today, partly because the sharp rise in hardware costs, with DRAM prices up more than 200% since early 2025, discourages buying ahead of need.
Purchases are tracking consumption, not fear.
The honest caveat
None of this means the cycle lasts forever.
There will almost certainly be a right-sizing of AI datacentre requirements eventually, with some cyclical fallout, as with every technology cycle.
But calling the timing is genuinely hard, particularly with the benefits of inference workloads only now beginning to accrue.
The Journal is right that these deals are no sure thing.
It is wrong to assume the memory industry has learned nothing from the last time it signed them.
