The most important line in this press release was buried near the bottom

Read the Commerce Department's July 29 announcement top to bottom and your eye goes exactly where the writers wanted it to go. Seven letters of intent signed under the CHIPS Act. $874 million total. GlobalFoundries at the top with $300 million, followed by a run of companies most readers have never heard of pulling down anywhere from $30 million to $245 million apiece. Photonics, new memory architectures, advanced packaging. Standard industrial-policy fare.

Then you get to one sentence near the end: "The Department will receive a minority, non-controlling equity stake in each company as a condition for receiving the funds to enhance the return for the U.S. taxpayer." Each company. All seven, no exceptions. That single clause converts the whole announcement from a subsidy story into an investment story, and it means the U.S. government just added itself to the cap table of seven semiconductor companies on the same afternoon.

GlobalFoundries went ahead and put a number on it. In its own release, the company stated that Commerce will receive equity "representing approximately 1% ownership of GF as of today's date." A Nasdaq-listed foundry disclosing, in an investor-facing document, that it is issuing shares to a federal agency. GF's market cap has been running roughly in the $27–32 billion range through mid-July, so $300 million landing near 1% checks out arithmetically.

The smaller companies were even blunter about the mechanics. Aeluma, also Nasdaq-listed, wrote that it "would issue equity securities to the U.S. Department of Commerce with a value equal to the total funding amount." That is $30 million in, $30 million of stock out. And the cash does not arrive in one lump — part comes up front, the rest is tied to milestones keyed to technical progress and eligible project costs.

So this is not a "Washington throws more money at chips" story. It is a story about how the 2022 CHIPS Act — designed around tax dollars going out as grants so companies would build fabs — has mutated by 2026 into something that behaves like a sovereign venture fund. Here is the deal: these seven letters are the clearest signal yet that equity-for-funding is no longer the exception. It is the default.

One foundry, six names you have never heard, and the office writing the checks

Start with the entity holding the checkbook. The announcement came from the CHIPS Research and Development Office (CRDO), housed under Commerce, and was published through NIST. Commerce Secretary Howard Lutnick framed it in political terms: "With today's compute supply chain investments, the Trump Administration is accelerating America's innovation engine." The working-level framing came from Bill Frauenhofer, the executive director for semiconductor innovation and investment, and it was far more technical — accelerating R&D across domestic photonics, novel memory architectures, and advanced packaging so that industry gets the extreme bandwidth and energy efficiency that complex AI workloads demand.

The headline recipient is GlobalFoundries. Most people file GF under "mature-node foundry for automotive and communications," which is fair, but it has been accumulating silicon photonics assets for over a decade. CEO Tim Breen put it this way: silicon photonics is essential to AI infrastructure, and while the industry has spent ten years describing the copper-to-light transition as something that is coming, it has now arrived. CTO Gregg Bartlett pointed to GF's already-qualified photonic device portfolio, 3D hybrid bonding, and advanced packaging capability as the basis for taking near-package optics (NPO) and co-packaged optics (CPO) all the way to volume production. The work lands at GF's fabs in Malta, New York and Burlington, Vermont.

The other six are, for most readers, complete unknowns. Kepler is building a new class of high-performance AI memory using 3D structures and ferroelectric technology, and at $245 million it is the second-largest award in the round. Multibeam is a privately held company in Sunnyvale, California that makes multi-column electron-beam lithography (MEBL) equipment; it was founded by Dr. David K. Lam, a genuine elder statesman of the semiconductor equipment business, and president Ken MacWilliams claims the new MBX platform cuts chip-to-chip power by more than 10x while compressing development cycles.

The strangest name on the list is Extropic. Led by Guillaume Verdon, the company is building what it calls a thermodynamic sampling unit (TSU) — an entirely new category of chip. The core idea is to treat the thermal fluctuations that occur naturally in ordinary CMOS transistors not as noise to be suppressed but as a computational resource, sampling directly from probability distributions. It is purpose-built for probabilistic AI workloads, and the company claims energy-efficiency gains of "orders of magnitude" over conventional GPUs on generative tasks. The remaining three — Thintronics, OBSIDIA Semiconductors, and Aeluma — cover ultra-low-loss interlayer dielectrics, non-destructive detection of counterfeit and malicious parts, and indium-phosphide-free large-diameter photonic substrates, respectively.

Look at the shape of that list. No Intel. No TSMC. No Samsung. Instead of the megacap manufacturers that soak up billions in facility capex, six of the seven are small or private companies holding technology that has not yet been proven at scale. That is the tell: CHIPS money is shifting its center of gravity from "cash to build a fab" to "cash to buy technology that does not exist yet."

One more piece of plumbing worth knowing. CRDO opened its Broad Agency Announcement (BAA number 2025-NIST-CHIPS-CRDO-01) on September 24, 2025. Minimum project budget of $10 million, maximum five-year period of performance, rolling submissions and rolling selections through September 2029. Buried in that BAA is the clause that separates it from every prior CHIPS program: applicants may be required to offer the government equity, warrants, intellectual property licenses, royalties, or revenue sharing. These seven deals are the first large-scale demonstration of what that clause looks like when it is actually enforced.

From $300 million down to $30 million — and what each one cost

Put the numbers in one table and the structure of the round jumps out. The top three take $685 million, or 78% of the total. The remaining four split $189 million.

Company Max award Technology focus Public/private
GlobalFoundries $300M Co-packaged optics, next-gen optical materials Nasdaq-listed
Kepler $245M 3D ferroelectric high-performance AI memory Private
Multibeam $140M Fine-pitch advanced packaging (multi-column e-beam) Private
Extropic $75M Thermodynamic sampling units (TSU) Private
Thintronics $50M Ultra-low-loss interlayer dielectrics Private
OBSIDIA Semiconductors $34M Non-destructive counterfeit/malicious part detection Private
Aeluma $30M InP-free large-diameter photonic substrates Nasdaq-listed
Total $874M — —

Why is the GF deal the one everyone points to? Because of what Commerce explicitly claimed it buys: the award will accelerate domestic co-packaged optics R&D "by 2–3 years." CPO means bolting the optical engine directly alongside the AI processor so you route around the power and bandwidth ceiling of copper traces — which happens to be the hottest bottleneck in AI data centers right now. GF's own materials list next-generation silicon photonics wafer technology, novel optical materials, 3D hybrid bonding, NPO/CPO, and its SCALE platform as being in scope.

Disclosure quality on the equity terms varies wildly from company to company. GF gave a percentage: approximately 1% ownership. Aeluma gave a mechanism: equity securities equal in value to the total funding. Multibeam's release specifies neither the equity structure nor the milestone structure. Commerce's own announcement contains no per-company ownership figures at all. Which means that as of today, nobody outside the negotiating rooms can add up exactly how much of these seven companies $874 million bought.

And here is the caveat that most coverage will skip: these are letters of intent, not contracts. Aeluma's language is worth quoting because it is the most explicit — the funding is subject to "additional due diligence, required approvals including internal U.S. government approvals, and the execution of definitive agreements by both parties." Extropic said the same thing in its own post, stating plainly that the arrangement is non-binding. How much of that $874 million ever actually moves is an open question.

The disbursement structure compounds this. Partial payment up front, remainder tied to technical milestones, is a sensible risk-management tool from the government's side. From a startup's side it means the announced number and the number that hits the bank account are two different things. In Extropic's case, standing up the first Z1 cluster and demonstrating performance on generative AI benchmarks is what unlocks the money. The company has laid out a roadmap running from the X0 prototype to Z1 to a Z1.5 built at a U.S. foundry — though it has not said which foundry.

Who actually walks away with something: the state, the companies, the taxpayer

Commerce got three things here. First, a technology portfolio. Optical interconnect, ferroelectric memory, e-beam packaging, probabilistic computing — one option purchased at each plausible chokepoint in the AI compute supply chain. Second, a political narrative. Lutnick has argued consistently since taking office that the government should stop handing out free grants and start capturing returns for taxpayers; equity conditions are that argument made operational. Third, speed. Because the BAA runs on rolling submission and rolling selection, CRDO can deploy capital far faster than the grinding due-diligence cycle of the original CHIPS incentives program.

On the company side, the value depends enormously on scale. For GF, $300 million is a modest single-digit percentage of annual revenue — not financially decisive. What GF is actually buying is time. Pulling CPO development forward two to three years means arriving in a market where Nvidia and Broadcom are already shipping, rather than showing up late. Against that, 1% dilution looks like a reasonable price.

For the small companies it is a different universe. At Aeluma's market cap, $30 million is existential money, and the equity cost is correspondingly heavy. CEO Jonathan Klamkin framed the goal as accelerating the development and commercialization of a scalable, high-performance semiconductor platform — which read the other way means that pace is not achievable without this check. For pre-revenue companies like Extropic or Kepler, federal money is not just cash; it is a validation stamp that says the U.S. government underwrote this technology. That flows directly into the valuation of the next private round.

The taxpayer's side of the ledger is the hardest to compute. In theory, if these companies succeed, the equity appreciates and the Treasury eventually captures the upside. In practice, most of these seven are early-stage firms whose core technology has not been demonstrated at production scale. Apply venture math and you expect most to go to zero while one or two carry the portfolio — except a venture fund can cut losers and defend its position with follow-on capital. A government agency cannot easily do either. Writing off a politically visible failed investment is dramatically harder than it is at a private fund.

There is a real skeptical case here and it deserves airtime. Ever since the Intel stake in 2025, critics have pointed out that a government which is simultaneously regulator and shareholder has a structural conflict of interest — an incentive to favor its own portfolio companies in permitting, procurement, or export decisions. Former Treasury Secretary Larry Summers described the Intel deal as "deal-based capitalism" rather than rule-based capitalism. Free-market institutions like the Cato Institute have used the blunter phrase: state capitalism. The genuinely interesting part is that this criticism arrives from both the left and the right at once.

SEMATECH worked, Solyndra went bankrupt: the two faces of government bets

The canonical American success story for direct federal money in semiconductors is SEMATECH. Founded in 1987, matched by DARPA at roughly $100 million per year for five years, and widely credited with reviving a U.S. equipment and materials ecosystem that was losing badly to Japan. The design detail that mattered: the government demanded no equity and no control. It funded shared research infrastructure and let member companies stay fierce competitors on everything above the base layer.

On the other side sits Solyndra. The Department of Energy issued a $535 million loan guarantee in 2009; the company filed for bankruptcy in 2011. The government had bet on one specific technology path — cylindrical CIGS thin film — and the premise collapsed when conventional silicon panel prices cratered. What often gets left out is that the same DOE program lent Tesla $465 million in 2010, and Tesla repaid it early in 2013. Same agency, same era, opposite outcomes. That is the whole nature of state-directed bets in two data points.

For direct equity specifically, the reference case is General Motors from 2008 to 2013. Treasury put in roughly $49.5 billion and at one point held 61% of the company as its largest shareholder. It sold the last of the stake in 2013 and locked in a loss north of $10 billion. Whether that counts as failure is still argued today, because the counterfactual involves several hundred thousand jobs. The AIG rescue ran the other way — Treasury exited that one with a profit.

The CHIPS program already has a failure of its own, and it is recent. Natcast, the nonprofit stood up to operate the National Semiconductor Technology Center (NSTC), was effectively dismantled in August 2025 when Commerce voided up to $7.4 billion in associated funding and terminated the agreement. NSTC operations reverted to direct NIST control, and planned R&D facilities in New York, California, and Arizona lost their budgets. Consider that of the $1.4 billion in first-round final awards under the National Advanced Packaging Manufacturing Program (NAPMP) announced in January 2025, $1.1 billion was earmarked for Natcast's packaging piloting facility — and this $874 million round starts to look partly like filling that hole with direct-to-company funding instead. For the record, the other NAPMP awards from that same round, including $100 million each to Absolics in Covington, Georgia for glass substrates and to Arizona State University, survived.

Nvidia is already shipping, Broadcom is already in volume

Zoom out to the CPO market and the $300 million looks a lot less commanding. Nvidia began shipping its new Spectrum-X CPO switches — co-developed with TSMC — to select partners in the first half of 2026, with up to 400 Tb/s of throughput and capacity expansion slated for the second half of the year. Broadcom's 51.2T Bailly CPO switch went into volume production alongside Delta Electronics and Micas Networks. TrendForce noted in a late-July report that the binding constraint has shifted away from the technology itself toward optical engine yield and advanced packaging capacity.

So the exact position GF is trying to reach two or three years early is already occupied by two enormous incumbents. Worse, both of them are riding TSMC's COUPE platform and SoIC-X 3D packaging, which means TSMC controls the practical gateway to optical engine integration. GF's edge is not going to come from winning a leading-edge node race it is not running. It comes from offering a fully domestic supply path out of Malta and Burlington — geopolitical differentiation, not process differentiation. That is why Breen leaned so hard on the phrase "in the United States."

For Korean and Taiwanese companies there are two distinct effects. The first is straightforward competition. The $458 million CHIPS award SK hynix finalized in December 2024 funds an HBM advanced packaging facility in West Lafayette, Indiana. The fine-pitch heterogeneous integration Multibeam is targeting with its $140 million, and GF's 3D hybrid bonding work, eventually collide with that same technical territory. Kepler's ferroelectric AI memory is an even more direct bet on what comes after HBM — and its $245 million exceeds half of SK hynix's entire finalized award.

The second effect is asymmetry in terms. Samsung Electronics finalized up to $4.745 billion in December 2024 with no equity condition attached. Same for TSMC at $6.6 billion and Micron in the $6.1 billion range. When reports surfaced in August 2025 that Lutnick might seek equity from those companies too, Taiwanese and Korean markets wobbled — until a government official drew a line, saying the administration was not looking to take stakes in TSMC or Micron. The reasoning was that companies which had already massively expanded U.S. investment were effectively exempt. But note what that exemption is: a judgment call, not a statute. Any new R&D money flowing through the BAA carries equity, warrants, and royalties as standard options, so a Korean firm reaching for fresh CHIPS R&D dollars should expect dilution to be on the table.

Watch the counterplay from other governments, too. Korea's approach under its K-Chips legislation runs through expanded tax credits and in-kind infrastructure support, and Samsung and SK hynix jointly announced large domestic semiconductor hub investments in June 2026. A tax credit, by construction, takes no ownership — philosophically the opposite of the American model. Over the next several years we are going to accumulate real comparative data on "grant plus equity" versus "tax credit plus infrastructure." It is too early to call a winner, but this is shaping up to be a genuinely useful natural experiment.

What this actually changes, depending on who you are

If you build AI infrastructure — nothing changes this quarter. Even accelerated, GF's CPO is years from volume, and what you will actually be racking in the meantime is Nvidia Spectrum-X and Broadcom Bailly. The directional signal is unambiguous though: the stretch of the rack where copper survives keeps shrinking, and you should be designing on the assumption that light comes all the way into the package. Practically, weight the interconnect line higher in your power budget models than you do today.

If you invest — do not push the headline number straight into a valuation model. These are letters of intent with due diligence, government approvals, and definitive agreements still pending, and the money is split between up-front and milestone tranches. On top of that, when a company issues new shares equal in value to the entire award (Aeluma's disclosed structure), the dilution is immediate and real. In small caps, the "secured federal funding" headline and the actual per-share value impact can point in opposite directions. For a large cap like GF, 1% dilution is noise; for a company whose award is a large fraction of its market cap, that calculation is the whole trade.

If you work in semis or manufacturing and might apply — the CRDO BAA is open through September 2029 with rolling submissions. Minimum project budget $10 million, period of performance up to five years. Applicants must be U.S. entities; for-profits, nonprofits, universities, FFRDCs, and federal agencies are all eligible. Before you file anything, read the clause on equity, warrants, IP licenses, royalties, and revenue sharing. When you build your financing comparison table, this belongs in the same column as a venture round, not in a separate "free money" bucket. The real decision variable is effective cost of capital versus an unconditional tax credit.

If you follow policy — through the first half of 2026, the count and dollar value of U.S. government equity positions across semiconductors, rare earths, lithium, quantum computing, and defense have kept climbing. These seven deals extend that line, and they use verbatim the same equity language as the May 21 announcement of $2.013 billion across nine quantum computing companies (IBM at $1 billion, GF at $375 million, among others). That is close to $3 billion out the door under an identical template in roughly two months. Whether this is a temporary experiment or a hardening institutional practice will be answered by the terms in the definitive agreements, not the press releases.

If you are just a person who uses this stuff — the honest answer is that you will feel this years from now, if at all. What is worth filing away is that every technology in this round shares one objective: more AI compute per watt. With data center power consumption spilling into electricity-bill politics and local infrastructure fights, optical interconnect and thermodynamic computing are attempts to solve that problem down at the physics layer. Whether or not they work, the fact that national budget is moving in this direction is itself information.

🥄 Three Things You’re Probably Wondering

— So what does this mean for me? If you are not directly invested, honestly, very little right now. But if you hold shares in a listed name like GF or Aeluma, the dilution terms are worth reading, and the broader drift of U.S. semiconductor policy from "grant" toward "equity investment" will shape the terms foreign chipmakers get on American soil.

— Is all $874 million actually going out the door? Not settled. All seven are letters of intent, with due diligence, internal government approvals, and definitive agreements still ahead. Payment is structured as partial up-front plus milestone-linked tranches, so a company that misses its technical targets does not collect the back end. The gap between announced and disbursed will not be visible for years.

— Is the government becoming a shareholder good or bad? Too early to call. The argument that taxpayers should share in the upside they funded is coherent. So is the concern about a regulator that is also an equity holder. The Intel stake from August 2025 — $8.9 billion for 10% — has not produced a clear verdict yet, and this model needs several more years before anyone can grade it.

Further Reading

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!