The Memory Gap Runs Through 2027

Most of the argument about chips and Taiwan is an argument about logic — the processors that do the computing. That is where the concentration is most extreme and where the coverage has settled. It is not where most American companies will feel a crisis first.

They will feel it in memory. Every server, every phone, every car, every AI accelerator needs DRAM, and the AI build-out needs a particular kind of it in volumes that did not exist three years ago. Memory is the part of the bill of materials that buyers historically treated as a commodity: interchangeable, always available, priced by the market rather than by geography. That assumption is the exposure.

The number the policy is built on

The United States currently produces less than 2 percent of the world’s advanced memory. The federal goal, set out in the CHIPS award to Micron for its Idaho project, is to raise that to roughly 10 percent by 2035.

Read those two figures next to each other. The stated national objective, fully funded and under construction, is to reach one tenth of global advanced memory production nine years from now. Even complete success leaves ninety percent of the world’s advanced memory made somewhere else, and most of that somewhere is in a band of East Asia that runs from South Korea through Taiwan.

Onshoring is real. It is also slow, and the calendar is the whole problem.

The construction is not symbolic. Micron is the only American-headquartered memory manufacturer, it is building two leading-edge fabs in Boise with about 600,000 square feet of cleanroom each, and in June 2025 it committed to roughly $200 billion in United States investment across Idaho, New York and Virginia, with a stated ambition of making 40 percent of its own DRAM in the United States. First output from the first Idaho fab is expected in 2027.

That is a serious industrial program. It is also a program whose earliest meaningful production arrives after the period in which planners are most worried about the Taiwan Strait, and whose full effect arrives in the 2030s. For anyone building a continuity plan today, the honest planning assumption is that the domestic capacity now under construction does not help in the scenario it is meant to address.

A fab is not a supply chain

There is a second timing problem, and it gets less attention because it involves a step most people outside the industry do not know exists.

High bandwidth memory — the stacked DRAM that sits beside an AI accelerator and largely determines how fast it runs — is not finished when the wafer is. It has to be stacked, packaged and tested, and that step is a specialized industry of its own, concentrated in Taiwan, South Korea, Japan and Singapore. Micron packages HBM in Taichung, Taiwan. Its plan brings advanced HBM packaging to the United States — at the Manassas, Virginia site — only after the Idaho fabs are producing wafers in volume.

So the sequence is: wafers first, packaging later. A company that reads the headlines and concludes that American memory production solves its exposure has missed the step where the most strategically important product in the category actually gets made. Wafers produced in Idaho in 2027 still finish their journey in Asia.

Micron itself is not pretending otherwise. In January 2026 it agreed to buy a fab in Miaoli County, Taiwan, for $1.8 billion, with meaningful wafer output expected in the second half of 2027. Building in Idaho and buying in Taiwan are not contradictory decisions. They are what a company does when demand is rising faster than any single geography can serve, and they are a reminder that the industry’s center of gravity has not moved yet.

The memory maker has already been a target

There is a counterintelligence dimension here that supply chain analysis usually leaves out.

On 21 May 2023, China’s Cyberspace Administration announced that Micron had failed a cybersecurity review and directed operators of critical information infrastructure in China to stop buying its products. The stated reason was security. The timing — months after Washington’s October 2022 export controls — pointed at retaliation, and analysts read it as a move calculated to pressure the United States while creating friction with South Korea, whose manufacturers stood to absorb the business.

Two things follow from that episode. The first is that market access for an American memory manufacturer is now a lever another government has already pulled, which makes it a lever that can be pulled again. The second matters more for anyone running security at one of these companies: a firm that a state has publicly designated as a problem is a firm whose people, processes and intellectual property are of standing interest to that state’s services. Memory process technology is among the most valuable intellectual property in manufacturing, and the industry’s history of trade secret cases involving DRAM is not short.

A company in that position is running two security problems at once: physical and operational risk to a multi-billion-dollar construction program, and a sustained foreign intelligence interest in what happens inside it. Those are not the same discipline, and organizations routinely staff only the first.

What the gap asks of a buyer

For the companies that buy memory rather than make it, the practical questions are narrower than the geopolitics suggest:

Know where the finishing happens, not just the brand on the part. Ask suppliers where wafers are fabricated, where packaging and test occur, and what the qualification timeline is for a second source. Most buyers can answer the first and not the second two.

Qualify alternatives before you need them. Re-qualifying memory in a regulated or safety-critical product is a program with a cost and a schedule, not a purchase order. Starting that work during a crisis means starting it behind every competitor who started earlier.

Plan for closure, not invasion. A quarantine, an inspection regime or an insurance market that stops writing cover interrupts shipments without a shot being fired. The threshold that disrupts supply is far below the threshold that dominates the public conversation.

Treat price signals as early warning. Memory is cyclical, so a price spike is usually just a cycle. But the AI build-out has tightened supply to the point where a modest geopolitical shock lands on a market with no slack in it, and contracts written against the old assumption of abundance are where the pain shows up first.

The point

American memory manufacturing is being rebuilt, seriously and at scale, with a public number attached to it. The number says one tenth of global supply, nine years out. Every milestone that would change an organization’s actual exposure — first Idaho output, domestic HBM packaging, a second source qualified into a product line — sits on the far side of the window people are planning for.

That is not an argument against onshoring. It is an argument for being honest about what it covers and when. The bet is already placed. The only question is whether the people who placed it know the settlement date.