For the better part of a decade, the world’s most advanced AI chips followed a single path to production: through Taiwan Semiconductor Manufacturing Co. Nvidia, AMD, Broadcom, and virtually every other fabless designer sent their most complex silicon to TSMC because no one else could match its yields at cutting-edge nodes. That calculus is starting to shift. Driven by billions in U.S. federal subsidies, lingering trauma from the 2021 chip shortage, and genuine progress on next-generation process technology, both Intel and Samsung are pulling in foundry interest from AI-chip designers who have decided that depending on a single supplier is a risk they can no longer afford.
The federal money that changed the math
The financial scaffolding behind this shift is now a matter of public record. In March 2024, the Commerce Department announced preliminary terms with Intel for up to $8.5 billion in direct CHIPS Act funding to expand domestic manufacturing. That figure was later revised to $7.86 billion in a March 2025 amendment, but the core commitment held: Intel would build and equip advanced fabs in Arizona, Ohio, Oregon, and New Mexico, with enforceable spending requirements documented in a binding agreement filed with the SEC in November 2024.
Samsung secured its own package in December 2024. The Commerce Department finalized up to $4.745 billion in direct funding for the South Korean manufacturer to expand leading-edge production in Taylor, Texas, organized around a cluster strategy covering logic, memory, and advanced packaging in a single geography.
Combined with the $6.6 billion awarded to TSMC for its Arizona fabs, the U.S. government has committed more than $19 billion in direct CHIPS Act funding to the three companies capable of manufacturing at the most advanced process nodes. The policy intent is explicit: build enough domestic capacity so that no single point of failure, whether geopolitical, logistical, or demand-driven, can paralyze the AI hardware supply chain the way the 2021 shortage paralyzed automotive and consumer electronics. That shortage cost the global auto industry alone an estimated $210 billion in lost revenue, according to consulting firm AlixPartners.
What Intel and Samsung are actually offering
Federal dollars create capacity, but winning foundry customers requires competitive technology. Both Intel and Samsung have been racing to close the gap with TSMC at the nodes that matter most for AI accelerators.
Intel’s foundry pitch centers on its 18A process node, which uses gate-all-around (GAA) transistors and backside power delivery, a combination TSMC is not expected to deploy together until its A14 node. Intel has said 18A is in risk production, with external customer tape-outs underway. The company’s foundry services division, rebranded as Intel Foundry, has publicly disclosed design-collaboration agreements with companies including Microsoft, though specific AI-chip tape-out details remain under wraps. The SEC-filed covenants tie Intel’s subsidy disbursements to concrete facility milestones, giving prospective customers unusual visibility into whether the capacity will actually materialize.
Samsung is betting on its own second-generation 2nm GAA process for advanced logic. The company has struggled publicly with yields at its 3nm GAA node, a problem widely covered in industry press and acknowledged indirectly by Samsung’s own delayed production ramp. The CHIPS Act funding is partly aimed at solving that problem by building a new, purpose-designed facility rather than retrofitting older lines. Whether the Taylor, Texas cluster can deliver yields competitive with TSMC’s N2 node remains one of the biggest open questions in the foundry market as of mid-2026.
Why designers are hedging now
The motivation for fabless companies to qualify a second source is not theoretical. During the 2021 shortage, TSMC’s allocation system forced even its largest customers to accept reduced wafer volumes, pushing product launches back by quarters. Companies that had no fallback had no leverage.
AI has made the stakes higher. Training clusters built around custom accelerators from companies like Google, Amazon, and Meta require massive wafer volumes at advanced nodes. Nvidia’s data-center GPU shipments have grown so rapidly that TSMC has reportedly dedicated entire fabs to its orders. For smaller AI-chip startups and hyperscalers designing custom silicon, the fear is straightforward: if demand for advanced wafers keeps climbing, TSMC may not have enough capacity for everyone, and those without a second qualified source will be last in line.
Qualifying a second foundry is expensive and slow. It typically requires 12 to 18 months of process development kit (PDK) work, test chip fabrication, and yield characterization before a designer can commit production volume. That timeline means companies making decisions in mid-2026 are already thinking about where they will manufacture chips shipping in 2028 and beyond. The CHIPS Act funding, by reducing the risk that Intel or Samsung will quietly shelve expansion plans if the market softens, makes it easier to justify that upfront investment.
What is still unproven
No named AI-chip designer has publicly confirmed shifting primary production volume from TSMC to Intel or Samsung. The federal funding packages prove that capacity is being built, but the question of whether fabless firms have committed high-volume orders to these new lines lacks direct evidence in available government or corporate filings as of June 2026.
Quantitative data on how many AI wafer starts are moving away from TSMC has not appeared in any public source. TSMC’s own capacity forecasts and customer allocation records for 2026 and 2027 production cycles have not been disclosed in a way that allows independent comparison. Without those numbers, the scale of any shift remains unclear, even if the direction is consistent with the policy intent behind both CHIPS Act awards.
Execution risk looms over every timeline. Advanced fabs routinely encounter construction delays, equipment bottlenecks, and qualification setbacks. Intel’s 18A ramp and Samsung’s 2nm GAA buildout both face the kind of technical hurdles that have historically pushed production schedules back by quarters. For chip designers planning products on 18- to 24-month cycles, a slip of even two quarters can determine whether a new foundry option is viable for a given generation of silicon.
Political durability adds another variable. Intel’s agreement is a binding legal document with SEC-filed covenants, and Samsung’s award has been finalized, but broader CHIPS Act implementation could face adjustments under future administrations. The disbursement and facility-completion timelines stretch well into the late 2020s. Policy continuity is assumed but not guaranteed.
Where this leaves the foundry race
The verified record, as of mid-2026, shows a U.S. government that has committed more than $19 billion to underwrite three competing advanced foundries on American soil. Intel and Samsung now have federally backed, contractually binding expansion plans that did not exist three years ago. TSMC, meanwhile, is building its own Arizona capacity, meaning the concentration risk that defined the 2021 shortage is being addressed from multiple directions simultaneously.
But infrastructure is not market share. Whether Intel and Samsung can convert subsidized capacity into a genuine competitive triopoly for AI chips depends on whether they deliver yields, turnaround times, and process maturity that fabless designers trust with their most valuable products. TSMC built its dominance over decades of relentless execution. Federal funding can accelerate the construction of fabs; it cannot accelerate the accumulation of manufacturing know-how.
What has changed is the risk calculation for designers. The cost of qualifying a second source is real, but the cost of being locked into a single supplier during the next shortage is higher. With binding federal commitments now backing both Intel and Samsung, the option to diversify is more credible than it has been at any point in the modern foundry era. The question is no longer whether alternatives to TSMC will exist. It is whether they will be good enough, soon enough, for the chips that matter most.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
