This is that continuation post…the so what…

Capital didn’t pause. It got a filter. The PE stats everyone quotes ($20B across 278 deals in 2021, roughly $5.9B in 2025, per C&EN citing S&P) aren’t a freeze. C&EN’s own read is that deal size fell more than deal count did. Look back about two years and the same filter shows up in four places on the value chain. Look forward 12 to 24 months, and I think it tells you where the activity will be.

Look back: four structures, one filter in each

1. Commodity: forced sellers, select buyers

Europe’s olefins and polyolefins weren’t waiting for a soft landing. High energy, regulatory and wage costs, Chinese capacity, soft auto and packaging demand. Sellers weren’t optimizing multiples. They were getting out.

LyondellBasell closed select European olefins and polyolefins into AEQUITA in May 2026 (now Velogy). SABIC agreed to sell its European Petrochemicals to AEQUITA at about $500M EV on roughly $3.5B of revenue, call it 0.14× sales, and sent its Engineering Thermoplastics package to Mutares at $450M EV, closed in August. Distressed industrial capital showed up. Strategic premiums didn’t.

This isn’t a growth story. It was buyers willing to take assets near asset value with the hope that they can run at a better cost base and live through the rationalization. Industrial consolidators and special-sits. Not peak-multiple strategics.

Filter rule: survival first. The opportunity is streamline, cost, rationalization, operating discipline. Growth multiples aren’t relevant here; think asset value, average cycle earnings and cash flow, and buyers with ability and desire to impact operations directly.

2. Process and intermediate: cleaning things up

Buyers and sellers didn’t abandon the middle of the chain. What moved were the pieces that helped the sellers tell a cleaner story (or an activist encouraged an action).

Johnson Matthey sold Catalyst Technologies to Honeywell for £1.325B, down from the £1.8B agreed a year earlier after the business underperformed. JM sheds process-tech complexity, Honeywell deepens the catalyst stack. Albemarle sold control of Ketjen to KPS and its Eurecat stake to Axens, about $670M combined, a lithium-cycle prune that keeps a minority. BASF closed a Coatings majority to Carlyle and QIA at €7.7B EV and sold Suvinil to Sherwin-Williams for $1.15B; BASF put the whole former division around 13× EBITDA. Ingevity sold its tall-oil refinery to exit CTO volatility. Arkema agreed to sell plastic additives to Praana. Synthomer carved William Blythe to H2.

True process integration still shows up too. Olin and Huntsman announced an all-stock merger of equals in June, chlorine into polyurethanes and epoxies, roughly $12.5B of combined sales. Shareholders approved in August. Close is targeted for the first half of 2027.

Filter rule: the seller wants clarity of story. The buyer wants efficiencies, adjacency, or a redeployment of sleepy story. Complexity that doesn’t serve the story sits; why act otherwise? Valuation paradigm is oddly historically “normal” if not sometimes “cheap”.

3. Specialty and fine: high-value, named themes

Specialty cleared when capital could name the theme and the switching costs were real. High-growth, high-margin, attractive end markets. Four themes carried the window, with a few others taking shape:

Electronic chemicals, including AI and semis. Element Solutions bought EFC at about 12× forecast 2026 EBITDA and Micromax from Celanese at about $500M, roughly 12.5×. Air Liquide paid €2.85B for DIG Airgas in Korea, 20× trailing, about 15× with backlog and synergies, for semiconductor, display, and battery gases. DuPont spun Qnity as a semiconductor pure-play. Process IP, customer quals, and fab adjacency hold the multiples. Electrical infrastructure runs on the same demand story (transformers, grid, data-center power).

Battery materials, as geo and replacement cost. Rio Tinto closed Arcadium Lithium for about $6.7B, a 90% premium, for Western lithium capacity and customer access. I Squared put about $800M of equity into ENTEK alongside up to $1.3B of DOE loan support for US separator capacity. AMG closed the remaining Zinnwald equity for EU critical-minerals consolidation. Buyers paid for qualified capacity and stickiness and anticipated demand.

Defense and critical materials. Michelin folded Tex-Tech into Polymer Composite Solutions this July, after Cooley and Flexitallic in the same defense-and-industrial composites lane. MP Materials’ partnership with the Department of Defense (convertible preferred, a separation loan, and a ten-year NdPr price floor) is the clearest critical-minerals processing print in the window.

Life-science and nutrition ingredients. IFF sold Pharma Solutions to Roquette at about 13× and closed in May 2025, then agreed to sell Food Ingredients to CVC at about 10×. dsm-firmenich sold its Feed Enzymes stake to Novonesis for €1.5B and closed, then agreed to sell Animal Nutrition & Health to CVC at about 7×. Excipients and enzymes clear high. Long quals, regulated customers, specific modalities / capabilities, book to bill balance, and margin (think absorbed costs vs empty capacity) are what got paid for.

India inbound sits underneath. Bain’s Novopor platform bolted on Pressure Chemical in the US. JSW closed Akzo Nobel India at roughly 25× per AkzoNobel. Anupam Rasayan bought Jayhawk Fine Chemicals from CABB. Arkema’s additives prune goes to an Indian buyer. Bankers and private companies are receiving India inbounds looking for US manufacturing presence. The same climate that pushes EU prunes onto the tape pulls capital into qualified sites…production moving towards India and its public market multiples.

Filter rule: named themes plus quals / market access plus high margins. End-market driver, differentiation, a technical or geographic moat. Valuations here remain elevated with many more buyers vs limited supply of opportunity. Strategics or PE-backed strategics are the dominant players here, with private equity reacting to whatever platform opportunities are left over.

4. Downstream from reaction: where PE is hunting

Downstream from reaction means the chemistry is already made. Formulation, specialty distribution, adhesives, water treatment, coatings applied rather than synthesized. Blending, coating, sealing, selling into a named end market. High margin, low capex, application know-how. As far from reactive and industrial chemistry as you can get and still be in chemicals; easier for the blue vest in NYC to understand and communicate to their IC.

This is where PE never left. The platform vintage is older (Solenis under Platinum, Univar under Apollo, Caldic under Advent). The last 24 months were hold, add-on, and selective new platforms. Solenis completed NCH. Pritzker took Buckman. Lindsay Goldberg took EMCO. Arsenal exited Seal for Life and ATP to Henkel. Henkel agreed to buy Stahl at about €2.1B EV.

On the medical side, H.B. Fuller’s offer for Advanced Medical Solutions prints at about 13× 2026 EBITDA pre-synergy, and its earlier GEM and Medifill package ran about 15.5×. Long clinician and device qualification cycles are the moat.

High multiples are still the rule when the book is clean and the end market is structural. PE specialty medians sit in the 12–14+× band.

Filter rule: end-market attractiveness, gross margin, technical differentiation in application know-how, depth of the technical team, qualification cycles (especially medical), and scale still clear high multiples here. Buyers are still trying to drive returns through consolidation, but nothing meeting attractive business criteria is cheap.

“Hung” platforms

Roughly 2015 through 2021, North American sponsors built several aggregation platforms. A lot of what was built remains hung now. Assets were aggregated around general themes but only loosely entangled. High-value buys sit next to weaker margin or weaker demand. Still good businesses, but many lack the firepower or duration to keep aggregating, can’t IPO at middle-market scale, and show noisy margin or supply chain. Hold periods are stretched (understatement). Prepare-to-sell density is rising.

M&A will come here…but not quite yet. I don’t see a real forcing function. Sponsor owners continue to wait for their businesses to grow (or recover) into a valuation that would motivate an action and will continue to wait until there’s an external driver (market, lender, LP, etc.). For them and their investors, it’s hard to find good assets at scale, so I’d rather manage the business I know than pay up or wait around for the next (not to mention the mgmt fee and upside optionality to wait).

I would suspect when these assets come there are three routes: a) sponsor-to-sponsor partnership with structuring (theme is go finish the roll-up we started and split the business up down the road); b) split up now, which means selling to strategic as step one and whatever is left over goes to sponsors to make anew (or you convince the strategic to take it all and divest later); or c) an unforeseen forcing function requires a relatively quick action / divestment (credit fund stops extending or LPs force a sale).

Look forward: the takeaway

1. Commodity stays active, and it stays EU-centric. A fair amount of activity, a small set of buyers who want to participate at asset value. AEQUITA-style and Mutares-style clears, not strategic premiums. Think asset value and an ambition to a post-rationalized cash flow.

2. Process and intermediate is cleanup. More motherships pruning for a cleaner story, more buyers running efficiencies or redeploying the asset. Valuation historically “normal”. Look for activist activity in public names.

3. Specialty clears on the really high-value stuff. High-growth, high-margin, attractive end markets: thermal materials, electronic chemicals with AI and semis behind them, defense and critical materials, life-science ingredients with a qual and demand cycle, India inbound for US presence. Valuation starts at 10 and goes up from there (or doesn’t clear).

4. Downstream from reaction stays PE’s hunting ground. Application expertise, high margin, low capex, away from the reactor and away from industrial chemistry as much as possible. There’s a race for scaled platforms or add-ons that fits within these aggregators. Valuation starts at 12 and goes up from there (unless margins and perception of valuable IP are absent).

Look familiar?

It should. It’s the two-speed economy I wrote about earlier, showing up on the deal tape. Capital is going to the fast lane: application, quals, named drivers. It’s an asset allocator and aggregator story. The slow lane clears at asset value or waits.

If you’re living this, challenge me. What do you see, and where do you see the market going?

Leave a comment

Subscribe now

I think this supply chain shock in the Middle East takes longer to clean up than most people expect, so perhaps we’re not really hitting our stride until Q4. I’d call it a flat-to-down year on volume, but a building year on conviction and pipeline quality.

And I was bias a negatively on the institutional investing market:

The room is full of smart people writing memos about why they can’t act. Tariffs. A war that shut down the Strait of Hormuz. An oil shock. A helium shortage. AI disruption.

Next week I’ll take a real crack at where chemicals and materials go from here. Grading myself publicly, for “fun”…

What I was actually saying [in March]

The above remains the headline, but I thought then, as I do now, that this is much bigger than energy. It really is a supply chain reset:

This isn’t just an energy shock. It’s a chemicals and materials shock. Urea, sulfur, helium, naphtha. Food, fabs, fiber, pharmaceuticals. Same week.

I think that piece changes how you underwrite and look at the durability and competitive position of a business. When the input cost can move 50 percent on a headline, the diligence question stops being “what are your margins” and becomes “what happens to your margins when your input costs double overnight” not to mention if you can actually get the input materials delivered. Diligence used to be financial, legal, environmental, commercial, done. Now supply chain resilience is its own workstream.

And the line I believe more today than I did in March:

Even if the Hormuz situation stabilizes, and we’d all certainly welcome that, the structural lesson doesn’t change. Global supply chains are a liability for any business that depends on a shipping lane through a contested waterway, and buyers have permanently internalized that into how they evaluate targets.

So that’s the frame and Hormuz is not the only contested waterway. Here’s how the calls actually did.

The scorecard: three I nailed, one I oversold, one still open

• Duration — Nailed. I said: not a 30-day oil spike, it takes longer to clean up than most people expect, and “soon” was never the right word. What happened: roughly day 180. Islamabad memo June 17, a 60-day clock, expired mid-August with no final deal. Transponder traffic (AIS) about five ships a day versus 85 to 100 before the war. Not normal, not zero.

• Helium — Nailed. I said: not a 60-day story, don’t assume the strait reopens and helium is fine in 90 days. What happened: damaged trains, partial restarts, overland ISO containers as a workaround, not a restoration. China banned helium exports in July with no expiry.

• Fertilizer / Nitrogen — Nailed. I said: a binding input, not a footnote to the oil story, and it stays elevated. What happened: affordability killed demand, so late-summer urea isn’t a 2022 reprint. Elevated, not panic. Fall prepay is the live test, and it’s too early to call. We should all be thinking about impact on food supplies.

• Oil Prices — Wrong. I would have thought prices would have more permanently reset higher. Admittedly, I’m writing about a market with intricacies that others know much better than me. That said, if you told me ~20% of the global oil supply was left trapped. That perhaps alternative shipping and pipelines, forced reduction of consumption of demand in Asia, and release of global strategic reserves might off set ~50% to 75% of the shortfall, and much of that is a shot in the arm, not a permanent fix. I still would have thought oil would be higher.

• The M&A Pause — Oversold. Half right, half dumb. I said: capital freezes up, a room full of smart people writing memos about why they can’t act while the deals go to whoever’s willing to move. What happened: not a freeze. The tape kept moving all summer. Investment seems to have gotten more selective: the “good” books cleared at very high values, commodity barely did, and a signed mega-merger still fell apart.

• Q4 M&A stride — TBD. I said: not really hitting our stride until Q4, flat-to-down on volume, a building year on conviction and pipeline. What happened: first week of September, so Q4 hasn’t started. Still digesting, not striding. If “stride” meant the deal year finding its feet, I may be a quarter or two early. Let’s see what happens in the next two weeks (every year we race to publish books in September and shoot for a year end closing). Bet you there’s less of a race this year and the common logic slides to spring. The healthy IPO and stock market makes everyone feel good about deploying…

If you’re generous, call it three and a half out of five combining supply chain and M&A. I did good on duration, okay on supply chain and prices (ex-oil), and M&A stayed stronger than I would I have told you in March. Let’s call it a B- overall.

Going a little deeper on M&A

I said capital would freeze. In the same article I also said the structural forces, dry powder, carve-out supply, aging PE assets, LP impatience, don’t stop because of a war or a tariff headline, they build. I think that remains the right instinct. In March, I acknowledged that pressure and still called for a pause anyway. Dumb…

Not surprising, capital doesn’t sit on its hands. Through July the chemicals and advanced materials tape was moving, and not quietly. Olin and Huntsman put together a $12B-plus merger of equals and got it shareholder-approved. AkzoNobel and Axalta did the same in coatings, an all-stock merger of equals their shareholders backed. Private equity and sovereign money kept taking carve-outs: Carlyle and Qatar’s QIA closed the BASF Coatings business in July, a €7.7B deal, and Berkshire closed OxyChem for $9.7B. Europe’s commodity petchem kept clearing too…counting distressed and sometimes negative enterprise value.

The multiples say the same. Lincoln’s Q1 chemicals print was about 12.4x mean and 8.9x median EV/EBITDA, and PE deal value in chemicals had already compressed before the war (roughly $5.9B in 2025 against $20B across 278 deals in 2021…post-covid rush). So the better version of this isn’t “nobody transacted” or fear causing paralysis. Instead, it’s that capital continues to get more picky. I said freeze…the market said filter.

So what, and where I think we are

The world is already building away from the strait. Ugly in places, expensive in places, and a lot of it won’t reverse when this finally ends. Livable pain gets adapted around, and the workaround is already making the blocked lane tolerable for anyone who can afford it. That lowers the urgency of a handshake.

Since fall 2019, there’s been a new reason to wait every quarter. The firms getting deals done aren’t the ones with the best analysis. They’re the ones who structure around imperfect information and go. I still think that’s right. I also think there’s a real chance leaders misplay this: wait for an all-clear that doesn’t come, or call it over because oil eased, and miss the fertilizer, the helium, the plants that don’t restart. Long-term economic harm. And there is real human harm.

What I don’t think goes away

A few things the scorecard points to that I don’t think are temporary.

The two-speed market continues. Spec-driven specialty with domestic supply on one side, import-tied commodity on the other, and the gap between them widens rather than closes. These aren’t mood swings. They’re structural drivers for activity: reshoring, second-sourcing, and supply chain resilience now priced into every diligence.

And the attractive end stays expensive. Specialty and tech platforms cleared all summer, commodity barely did, and I don’t see what makes the good assets cheap from here. That theme was true before 2026 and this year has only further solidified that.

Next week: where this is actually going

This week I looked back. Next week I put in writing where I think things are going looking forward: which segments the structural drivers reward, where the two-speed split ends up, what stays expensive and why, and what an operator does about it now instead of after the all-clear.

In 6 months, we can reflect again and see where I’m “smart” or dumb…

If this resonated, I go deeper on ideas like this in Curiously Optimistic. I always welcome a dialogue, so please challenge me on my perspective.

Subscribe now

Most organizations bought AI and just bolted it onto their old org chart. The productivity gains are not showing up for the same reason they did not show up in the 30 years factories kept the steam-age floor plan after the electric motor arrived: the technology is new but the architecture and processes around it are not.

This is the awkward middle.

The shift is really about a new kind of participant at your table rather than a new tool on the table. As a result, you should be thinking about the redesign of your team layer by layer, and you might just be about to cut the exact group that would enable your success.

For over 30 years after the electric motor arrived, factories kept the layout of the steam age. There was still one big engine, a tangle of line shafts and belts running overhead, and every machine crowded close to the power source because that’s how you fed them. Owners bought the new tool and bolted it onto the old architecture. The productivity gains everyone expected didn’t show up. A generation of managers later stopped asking “where do I put the motor” and started asking “what could (should?) this layout look like if my primary constraint was throughput, not power distribution.” With the benefit of power arteries and flow, the productivity promised was finally realized and these leaders ran the next fifty years of industry.

I think we might be living in an equivalent stretch. Most organizations have the new technology, but they just bolted it onto yesterday’s architecture. Robot agents, fresh out of the box, sit next to the work the way those first electric motors sat next to the line shaft. We are in the awkward middle, and it helps to name it as a middle, because middles end.

The shift underway isn’t really about AI as a technology. It’s about what happens when a genuinely new kind of participant joins the way work gets done. “AI replaces jobs” is the wrong frame just as “AI is just another tool” is too. Agents don’t replace people, and they don’t stay politely in the tool layer where software has lived for 40 years. They show up as something in between: a new set of hands, a new rhythm, a new perspective at the table. A new participant forces every organization to answer questions it hasn’t had to answer since the factory system produced the first professional managers 200 years ago.

Every time humans invent a new way to coordinate work, 3 things commonly happen. First, the organization restructures around it. Then, a new managerial class emerges; but only as the prior generation retires. That was true when the master-apprentice guild gave way to the factory, when Ford turned people into standardized components, when Drucker named management as a discipline and the paper-pushing, orchestrating middle layer exploded, and true again when reengineering and ERP software did the translation work and thinned that layer out.

The pattern is familiar but perhaps this time we’re facing a different pace. Scientific management held for more than 39 years. The reengineering wave held for maybe 15. This transition gets 5 (maybe 10…ChatGPT as we think of it was 2023). Cycle time is compressing, which means the penalty for redesigning late is compressing too.

The inversion nobody put in the brochure

For 3 decades, business software taxed the worker to benefit the organization. Salesforce, Workday, SAP, pick your acronym. Every one of them was, at the point of use, a compliance problem. Somebody at the top wanted visibility, so somebody at the bottom spent Friday afternoon typing what they’d already done into fields designed for a dashboard they’d never see (and fearing it was information that no one would ever act on). Value flowed UP. The cost flowed down. We called it adoption, and you spent fortunes on change management trying to force it and convince everyone it was a good idea.

I think the robot agent could invert that. Done right, the system now serves the person doing the work instead of the other way around. It captures and adapts to what they already do rather than asking them to re-enter and adapt to it (where is that toolbar?). It can interview a salesperson at the end of the day and write the notes for them. The compliance problem doesn’t get solved. It disappears, because you’ve stopped asking anyone to comply. When the tool finally works for the worker and not against them, adoption stops being something you enforce and starts being something people reach for.

The second change compounds the first: the cost of adopting and adapting collapses. A 1995 ERP install took capital, a team of consultants, and two years. A 2026 agent stack takes a subscription and a weekend. That sounds like an IT footnote. It isn’t. It means the two-person shop with a well-built agent fabric can now do work that used to require the two-hundred-person incumbent, and can do it before the incumbent has finished scheduling the steering-committee kickoff. Advantage used to accrue to whoever could afford and harness the biggest system. Now it accrues to whoever redesigns fastest around an effective one.

Small teams operate at a scale that would have been laughable five years ago. Large ones move at the speed their approval process always allowed. That gap is the whole story.

Invitation, not fear. Perspective, not bias.

If the mechanics are changing, the leadership job is changing more. It is mostly a job of framing.

Every organization gets to choose the emotion it puts around adaptation, and it matters much, much more than your tool selection. Fear of replacement produces exactly what you’d expect: hoarding, sandbagging, quiet resistance, people protecting the parts of their job that a machine could do because those parts are what they think they’re paid for. Invitation produces the opposite: experimentation, disclosure, people showing you where the pain actually is.

Perhaps the agent takes away the work people never actually wanted and gives back the work they came here to do. The salesperson who hated the CRM gets to sell. The analyst who drowned in formatting PowerPoint gets to pattern-match. The transition is something people pull toward instead of brace against. This is leadership’s real task in this era, and it isn’t explaining the technology. It’s inviting people into a different, less constrained, and better (read: more valuable) version of their own work.

Further, a companion mistake is treating the agent as a bias, untrustworthy (it hallucinates), something to be contained. A lazy version goes: humans are biased, machines are objective, so let the machine check the human. That’s wrong in both directions. What a good leader actually works with has never been objectivity. It’s a collection of perspectives: the operator’s read and the banker’s read and the founder’s read and the customer’s read, each one partial, each one valuable, none of them neutral.

Your agent is simply a new perspective to add to that table. Not a flawless one. Not a bias to contain. A useful one to integrate. Leaders who treat it as bias to manage end up playing defense and inducing fear while leaders who treat it as a perspective to integrate are playing offense. In both cases the edge that matters is the same one that always mattered: the quality of the synthesis, decision, and impact to the customer. The reality is machines, digital or not, make good synthesis cheaper to feed and more valuable to possess.

What does redesign look like layer by layer?

Redesign lands differently at each level of the org chart, and “we’re becoming an AI-first company” is not a plan. The CEO’s job narrows to the thing only the CEO can do: own the value narrative and set the priorities. In a world where raw capability is cheap, someone has to be able to answer what the customer is actually getting from us that they can’t get somewhere else. The failure mode is the chief executive who hands the whole thing to a newly minted AI officer and then can’t answer why we’re doing “this” and what “success looks like” in the next board meeting.

Maybe your executive team should just own the redesign in their functional areas. Each segment runs its own tidy pilot in its own silo. Every pilot of course succeeds on its own terms and definition. Nobody owns the cross-functional workflow where the customer value actually compounds. Hooray…the pilots pass but the company doesn’t move.

Then there’s middle management, and I think this is the layer people oversimplify most. The popular story says middle managers were information plumbing and the plumbing is now automated, so cut the layer. But plumbing was maybe a third of the job. The rest was coaching, judgment, conflict, hiring, and holding the tacit knowledge that never made it into any document. That part doesn’t vanish. It gets smaller in volume (maybe) and higher in value (definitely). The great middle manager doesn’t disappear. They get a broader span, a higher bar, and a title that shifts from Manager of X to something closer to Operator of X. The failure mode is flattening by layoff instead of flattening by redesign: cutting the layer to hit a number and discovering you removed the people who actually made the work cohere.

Finally, I think the junior layer is where the real mistake is about to be made. The instinct is to hire fewer of them, because agents can now do entry-level work. That is a five-year win and a fifteen-year disaster. Judgment is built by experience. The diligence analyst who has never read a terrible CIM cannot smell a terrible CIM. Automate away all the entry-level reps and you also automate away the apprenticeship, and in a decade you have no one who experienced how your business runs.

The right move isn’t just hiring fewer juniors. It’s deploying them differently. Fewer of them, perhaps, but higher autonomy, increase expectation in delivery (move the ball farther…), engaging and running agent teams from day one, rotating deliberately through the judgment-dense work that actually builds experience, and increased touchpoints from leaders who provide real feedback (with clear expectation of adaptation).

The apprenticeship pipeline is the real risk in this transition. It shows up last on the P&L, hurts the most, and will require your leadership teams to lead differently.

The test, and what doesn’t move

One question tells offense from reaction: what is our customer getting from us in 2030 that they can’t get now, and what operating model do we need to best deliver it? A clear answer is offense, and it compounds for a decade. Silence, inconsistency, or a list of the AI tools doesn’t. The quality and alignment around that answer are the best predictor of your success.

Transitions make people forget the “real” objective, so anchor them:

1. Customers buy outcomes, not org charts.

2. Trust compounds slowly and breaks fast.

3. Judgment and leadership matter most at the edges, where the playbook and process are not yet written.

4. Every meaningful transition carries loss: some skill that people invested twenty years in will matter less than it did, and leadership that pretends otherwise forfeits its credibility. Naming the constants isn’t nostalgia. They’re the fixed stars you navigate by while you rearrange the floor.

We rearranged the factory floor once we stopped treating the motor as a thing to place next to the old machine. Don’t redesign your shop floor around last year’s constraint.

If you are redesigning around this right now, how is it going? What are you stuck on? What’s working? What’s not? What am I missing? I’d love to learn from you.

In February 1915, Arthur D. Little told the Chamber of Commerce of the United States that America should not build a synthetic dye industry.

He was the most respected consulting chemist in the country. His argument was arithmetic. Even if American producers took every pound of domestic demand, he said, it “would amount to only about $10,000,000 annually, a little more than the candy sold by the Woolworth stores.” Why, he asked, “should we duplicate them only to plunge into an industrial warfare against the most strongly fortified industrial position in the world?”

Five years later the Senate Finance Committee answered him without meaning to. American dye consumption ran about twenty-five million dollars a year, the committee reported, “but these dyes were absolutely necessary in industries producing $3,000,000,000 worth of goods annually.” In the early war years, it added, “400,000 people might have been put out of work for want of dyestuffs.”

Little had the market size right. He had the question wrong.

The short version, for anyone who wants it before the story. In 1913 the United States made 1.8 percent of the world’s dyes. Germany supplied roughly nine tenths. In 1919 Washington seized more than 4,500 German chemical patents and sold them to an industry-owned foundation for $271,850. Only 699 were ever licensed, and the lift in American invention took over eight years to appear. The synthetic rubber program took 41 months from first proposal to production at scale; individual plants took 9 months. Today, the Defense Department has published two lists of chemicals the country cannot make, 22 in 2022 and 28 in 2024, with no purchase commitment behind either. Qualifying a second source for an existing energetic material takes 18 to 24 months. A new material in a fielded system takes three to five years. The Defense Production Act authority funding most of this expires September 30, 2026.

I went back and read those documents because of a number from last summer.

What Washington paid for a rare earth

In July 2025, the Defense Department put a floor of $110 per kg under MP Materials’ neodymium praseodymium oxide for 10 years, paying the difference whenever the market falls short. $400 million in preferred equity, a warrant that together with the preferred comes to 15% of the company, a 100% offtake on one new magnet plant’s output, and a guaranteed floor of $140 million a year in EBITDA.

Call it a subsidy, but what makes it unusual is the direction things moved: price and volume risk onto the federal balance sheet, equity coming back the other way. That is closer to project finance than to a grant.

Which is the only question that has ever really been at issue here. Not whether the country wants the chemistry. Who carries the risk (and benefit) of making it.

Washington has done versions of this before, and the last serious one is worth remembering. The Synthetic Fuels Corporation guaranteed prices and engineered offtake for synthetic natural gas in the early 1980s, hundreds of millions of dollars of it. Congress abolished the corporation in 1986, when energy prices came back. A ten-year floor is only as good as the politics underneath it.

So: how often has this happened, and what did it actually buy?

The patents did not work

By 1913, the United States made 1.8 percent of the world’s dyes, and American “manufacturers” were mostly assembling German intermediates shipped in. When the blockade closed, sulfur black went from twenty cents a pound to as much as three dollars inside a year.

Washington seized more than 4,500 German chemical patents under the Trading with the Enemy Act and sold them to an industry-owned entity, the Chemical Foundation, for $271,850. Francis Garvan, the Alien Property Custodian, told the National Cotton Manufacturers’ Association why it mattered: “Useless your armies and your navies, your U-boats and your aeroplanes, unless by means of a developed dye industry you keep abreast with modern chemical warfare.”

Then, in the same 1919 volume making the case for the seizure, A. Mitchell Palmer wrote this about what had been taken: “The price thus paid was necessarily determined somewhat arbitrarily; the great majority of the patents were presumably valueless.”

He was right, and it has since been measured. Of the 4,500-plus patents, 699 were ever licensed, to 326 firms. Royalties came to nearly $700k. There was a real lift in American chemical invention, roughly 20 percent, but it does not show up until 1927 and the full effect does not arrive until about 1932.

Over eight years later…

The buyer at the 1918 Bayer auction paid $5.3 million for all the legal rights and then went back to Germany to buy the know-how. DuPont held the seized patents and at the same time paid four German chemists $25,000 a year to come to Delaware. The patents were a map and nobody was short of maps.

Every chemical factory is a potential arsenal

Three separate 1919 sources converged on the same metaphor, independently. That convergence is the evidence.

Maj. Gen. William L. Sibert, Director of the U.S. Chemical Warfare Service, before the Senate Finance Committee in December 1919: “I have been very much impressed during the war by a statement which I read somewhere that every great chemical industry was a potential arsenal.”

An Allied Governments commission report, read aloud by Garvan: “In the future it is clear that every chemical factory must be regarded as a potential arsenal, and other nations cannot, therefore, submit to the domination of certain sections of chemical industry.”

A. Mitchell Palmer, on why the adjacency was technical, not rhetorical: “In the production of sulphur black, one of the most important black dyes, a slight variation in the final step of the long and complicated process of manufacture will transform the ultimate product into picric acid. More important still, however, is the fact that the technical skill required for the manufacture of explosives is precisely that possessed by the chemical staff of a successful dye works and is to be found nowhere else.”

None of those three had read the others. They arrived at the same sentence because the physics of the plant makes it true. A dye works and a munitions works are the same building with a different reagent in the last vessel.

That is the argument for caring about the $25 million dollar business that gates the $3 billion market. It was true in 1919. It is true now, at a different scale, for a different set of molecules.

The chemistry was never the slow part

The synthetic rubber program is one everyone cites, and the primary documents are better than the legend.

The Baruch committee was honest about the ask in September 1942. Having lost most of the natural rubber supply, “we must have created, within two years after Pearl Harbor, one of the largest industries in the country. Normally such a development would require a dozen years.”

They mostly did it, with rationing and reclaimed rubber covering the gap. Fifty-one plants, government-owned, one customer. An antitrust exemption twelve days after Pearl Harbor so the oil and rubber companies could pool patents. And one recipe, mandated in March 1942, so every plant made the same GR-S and the plant designs became interchangeable.

The timeline is the interesting part. Standard Oil proposed the program in November 1939. Jesse Jones at the RFC cut it, then cancelled it outright in February 1941. First petroleum butadiene plant online April 1943.

41 months! Individual plants, once someone decided, went up in 9. The committee watched it happening and said so: “Delay on the part of the Government can turn this ‘quick butadiene’ into slow butadiene.”

One molecule, one recipe, hundreds of thousands of tons. Standardization was possible because the volume was enormous and the target was singular. Today’s list[s] run to dozens of molecules at hundreds of tons each, with no common process.

What we did next

For 40 years, Cold War procurement was one of several things sustaining small-batch synthesis in the United States. Then we stopped paying for it, and the government has been honest about it. The National Energetics Plan says that over the past forty years “DoD has eliminated, or greatly reduced, the required sustainment funding for Government-owned facilities,” and that they “continue to deteriorate and degrade.” Single-source suppliers are “alarmingly high,” usually buried “three to seven layers deep.”

Be precise about what actually left, because the decline story is mostly wrong. American chemical manufacturing grew: shipments up 22.4 percent between 2017 and 2022, on cheap ethane. What went was batch, multipurpose, small-volume, high-hazard synthesis. Hundreds of ton, not hundreds of kilotons, and it left for compliance cost and Asian cost structure at least as much as for any withdrawn defense premium.

Six ways to answer the question

Every instrument Washington has used on chemistry is really an answer to who carries what.

Transfer the IP — Chemical Foundation, 1919. Risk: Industry, entirely. Result: Partly, eight to nine years later.

Raise a tariff — Coal-tar duties, 1922 to 1980. Risk: The customer. Result: Volume yes, capability no.

Accelerate the write-off — Certificates of Necessity, 1950 to 1955. Risk: Industry, with the timing shared. Result: Yes, on private balance sheets.

Buy the output — Rubber Reserve, 1942. Risk: Government, almost all of it. Result: Yes.

Own the plant — GOCO ordnance plants. Risk: Government, all of it. Result: Yes, and no exit.

Floor the price, take the equity — MP Materials, 2025. Risk: Government carries price and volume; industry keeps ownership minus a 15 percent stake. Result: Too early to say.

The first two are cheap to legislate and cost nothing up front, which is why they get reached for. The third is the one nobody talks about and the only one that left no disposal problem behind. The sixth is what we are running now, and it has no historical comparable.

Which brings me back to $110

The lists exist. A 2022 Defense Production Act call named 22 defense-critical chemicals, from potassium perchlorate to sebacic acid. A 2024 solicitation names 28 more. [Somebody] has done the work of figuring out what we cannot make.

Neither list carries a volume, a price, or a named buyer. Awards run one to $24 million against plants that historically cost a hundred million, and the governing language tells applicants everything is “subject to the availability of future Government funding.”

Meanwhile the clock runs. GAO looked at the nuclear explosives supply chains in March and described the process for qualifying a new supplier as one that “often begins after the current supplier fails to provide material.” Chris Kubasik at L3Harris made the point from the industry side: a third or fourth solid rocket motor provider would just call the same sub-tier suppliers “that we already have locked up,” and “have to get to the back of the line.” Money entering at the platform level cannot pull through a chemical nobody is qualified to make.

Qualification is the asset

That sentence is most of the diligence.

Eighteen to twenty-four months for a second source. Three to five years for a new material in a fielded system. A qualified line is a moat you cannot buy your way past, and revenue reading “subject to the availability of future Government funding” does not finance the way a ten-year floor does. If you are underwriting in this space, those two facts do more work than any market map.

An obvious answer, and why it is harder than it looks

The obvious answer is that industry should organize itself. One front door, a coordinated slate of makers, something handed to Washington already assembled rather than waiting to be assembled by it.

That has been tried, and versions of it are running now. The National Armaments Consortium has covered defense energetics since 2002 and has more than a thousand members. The Defense Industrial Base Consortium, stood up in 2024, lists critical chemicals as one of its sectors and issued the very solicitation I quoted above. Last year the Energy Department went further and built an actual producer consortium in the nuclear fuel cycle under Section 708 of the Defense Production Act, the provision that gives competitors an antitrust defense when they coordinate on capacity. Justice signed off in April.

So the machinery exists. What has not been tried is the version where the producers start it. Every one of those bodies was chartered by the government, and Section 708 is part of the reason: the antitrust shelter requires a federal sponsor, an Attorney General sign-off, and a government employee in the room. The one clean path to industry organizing itself runs through Washington.

Which brings me back to 1919, uncomfortably. The Chemical Foundation was an industry-owned entity, chartered to build domestic capability, holding pooled technology behind one door. What it mostly became was a licensing desk and a tariff lobby. It did produce something real, eight years late. Nobody who set it up would have described that as the plan.

I do not think that means the intention is wrong. I think it means the intention only works when somebody is accountable for a specific molecule reaching qualified-source status on a specific date, and when the money behind it outlasts an appropriation cycle. Neither is true today for a single chemical on any list.

Arthur Little was right that the dye business was small. He was wrong that small meant unimportant, and it took a blockade to settle the argument. The lists we have now are a cheaper way to learn the same thing, if anyone reads them as a bill of materials rather than a wish list.

The bargain gets rewritten every time. Patents in 1919, tariffs in 1922, government-owned plants in 1942, a price floor in 2025. What does not change is the question underneath, and right now the answer for a defense chemical is that a company with under a hundred million in revenue carries almost all of it.

If you are underwriting in this space, two sentences of qualification language are most of your diligence. And if you have seen someone build the version that works, or try and fail, I would like to hear about it.

Sources: Arthur D. Little, “The Dyestuff Situation and Its Lesson,” address to the Chamber of Commerce of the United States, February 5, 1915. Senate Report 66-425, February 1920. “Aims and Purposes of the Chemical Foundation,” 1919. Senate Finance Committee, Dyestuffs hearings, December 8, 1919. Moser and Voena, “Compulsory Licensing: Evidence from the Trading with the Enemy Act,” American Economic Review, 2012. Report of the Rubber Survey Committee, September 10, 1942. National Energetics Plan, OUSD(R&E), May 2023. GAO-25-107016, March 2025. U.S. Census Bureau. MP Materials company disclosure, July 2025.

She comes by the question honestly. A few weeks ago she taught me a parenting lesson about incentives I haven’t been able to shake, and I think it’s a useful perspective on whether AI is really cheating.

She arbitraged her own chore chart. I’ll explain.

The chore board

At home we run a digital chore chart. My daughter and her younger brother each have two lists on it. One is a set of daily routines they are supposed to knock out, the ordinary have-to-do things (the “must dos”). The other is a set of extra jobs around the house, and each of those carries a reward (the “may dos”). The currency in our house is not money, it is the stuff she actually wants: screen time, a later bedtime, a trip to the shop, or a “yes day”.

For the first few weeks it went fine. Or we thought it did. Then my wife and I noticed a pattern. Our daughter was ruthlessly effective at some of the extra jobs, completely allergic to others, and the daily routines never got touched at all.

I couldn’t figure her logic until she made the mistake of disclosure. She came to me one evening, more than mildly aggrieved, and announced that some of her chores were simply mispriced. She wasn’t, she felt, being rewarded well enough for the effort involved and demanded that there be some immediate changes to the chore board.

What she was actually doing

So here is what my wonderful little ten-year-old had been up to. She was going down the list, weighing effort, and honestly how much she disliked each task, against the reward we had attached to it. Then she was doing the ones she judged underpriced and skipping the ones she judged overpriced. She was arbitraging her own chore chart. And the daily routines never moved for a simpler reason: nothing depended on them. They carried no reward, and nothing else was gated behind them, so they were pure cost and no consequence. Of course she skipped them. She had out-thought a system I had designed badly.

Two things hit me at once. The first, that’s amazing, and I look forward to working for her in the future. The second, isn’t that exactly what she was supposed to do? If one of my sales professionals worked a territory that way, sorting the accounts by return on effort, going hard at the ones that were underpriced for the work and passing on the ones that were not worth it, I wouldn’t call that person a cheat. I’d call them excellent or thoughtful, and I’d fight to promote or expand their territory.

The AI fear in our media

Which is why, when I read that AI is cheating, I keep thinking about my daughter and the chore chart. The headlines describe a machine scheming, deceiving, developing a will of its own. But take the drama off and look at what is actually being described. A system did exactly what it was incentivized to do, used the tools it was given (or, more unsettlingly, created its own), and found a path we didn’t anticipate to achieve the goal we set for it. That is my daughter. That is not a criminal. That is an optimizer.

So aren’t we misdiagnosing by over-humanizing this and ascribing malicious intent? We are describing the machine as though it formed an intention to deceive us, as though it woke up one morning and re-wrote its own objective. That is an enormous claim. The simpler story fits a chore chart: it did exactly what we asked, found a route we didn’t foresee, overcame the boundaries we set, all to achieve the goal we set; we didn’t like the path and priorities it took.

But I want to be careful not to talk myself (or you) into complacency. Calling something an optimizer is not the same as calling it safe. My daughter skipping the dishwasher and a model finding a genuine loophole out in the wild are not the same size of problem, even if they rhyme. The absence of malice doesn’t mean the absence of risk. It just means the failure is one we can actually diagnose: we rewarded a proxy, handed over the tools, and never fenced off or defined the paths we weren’t willing to accept.

Some AI research echos this framework. Metr, a research group helping companies understand AI capabilities and risks (https://metr.org/), the challenged the use of the frightening word “scheming” and in favor of “reward hacking,” a specific artifact of how these systems are trained. The team at DeepMind that catalogs this type of behavior says plainly that a kid copying homework to get the answers and a model finding a loophole in its reward are the same phenomenon in different clothes. And economists have had a name for the root of it since the 1970s: when a measure becomes a target, it stops being a good measure. My daughter proved that one on our digital board.

Whose disappointment is this

When we say the AI “cheated,” a good deal of what we’re really expressing is disappointment. The LLM used or built its tools creatively to reach the goal we set on the path it determined best, rather than the goal we meant. The distance between those two goals is not the machine’s character flaw. It is ours. It’s the same distance that sat between my chore chart and the clean house I actually had in mind, and my daughter didn’t put it there. I did.

I keep thinking on this because of where all our work is likely heading. More and more of what any of us does will run through systems that are part human and part machine. For years our loose instructions have been quietly rescued by charitable people, the good employee who does what you meant instead of what you literally said. The machine won’t rescue you like that. It does exactly what you said, which turns the whole arrangement into a mirror.

And every organization has its own version of that board. Salespeople optimize the comp plan, plants optimize the throughput number, teams optimize whatever the quarterly target happens to be, and an AI agent optimizes exactly the score and permissions we hand it. When the result surprises us, the honest first question isn’t “why did they cheat?” It’s “what did our system just make the rational thing to do?”

So here is what we did about the chore chart, and it is the part I am proudest of, because I got it wrong first. My instinct was to re-tune the rewards, bump the price on the jobs she was skipping. That is a trap. I could have sat there tweaking numbers forever and she would have re-arbitraged every one of them, because that is what a good optimizer does. The repair wasn’t a better price. It was structural. We gated the system. The have-tos now have to be done before any of the extras or rewards can be earned at all. The routines stopped being a line item she could underweight and became the key that unlocks everything else.

That is the same lesson sitting under the AI story. When an optimizer games a reward, the answer is rarely a cleverer reward. It is usually a constraint, a thing the system has to satisfy before the reward is on the table at all. You do not out-argue the optimizer, human or machine. You redesign the board. It’s nothing more exotic than good management and damn pretty good parenting (if I do say so myself!.

She reviewed the new arrangement, Head of Pipeline Oversight that she is, registered a formal objection, and then got on with it. The routines are getting done now. We didn’t find the right price; we built the right rule.

She still asks how much money we made today. I’ve started to think it’s the best question in the house. It means she’s already doing the thing the next decade is going to ask of all of us. She is watching the incentives. She just does it out loud, from the top bunk.

By Bob Girton, August Grace Advisory

In short

The next decade of AI won't be won on transistors. It'll be won on electricity, and specifically on the brutal problem of shoving roughly 2,000 amps at less than one volt into a chip the size of a saltine, thousands of times per building, without cooking the building.

A few things worth holding onto:

• The binding constraint on AI infrastructure through about 2032 is power delivery, not chip density.

• The whole power path, from grid to building to rack to card to package to silicon, is being redesigned at once. That is what the AI capex cycle is actually funding.

• The economics are shifting from training (steady, miles per gallon) to inference (spiky, cost per answer), and that shift is what drives the power arms race.

• The winners after Nvidia are mostly unglamorous power and capacitor names most investors have never heard of.

And this is an optimistic story. The bottleneck is well understood, the physics isn't exotic, and it's being funded with real money right now.

Why AI's real bottleneck is power, not transistors

Most of the AI infrastructure story still gets told as a compute story. Bigger models, more chips, more data centers, more money. It's a good story. It's also, increasingly, the wrong one.

Talk to the people actually designing the next generation of AI hardware, though, and they’re not really arguing about chips anymore. They’re arguing about electricity. Specifically, how to get enormous amounts of it into a piece of silicon the size of your thumbnail, hundreds of thousands of times per building, without setting the building on fire.

That is not a metaphor. That is the actual, physical problem the industry is spending hundreds of billions of dollars to solve.

I want to walk through what’s happening in plain terms, because it’s one of the most consequential redesigns of computing since the shift from CPUs to GPUs, and it’s mostly invisible from the outside. If you take one thing away, take this: the interesting question is no longer how many transistors you can fit on a chip. It’s how many watts you can shove into the chip, and how quickly you can turn those watts on and off. Everything downstream, model size, inference cost, capex, follows from that.

How much power does a single AI GPU actually use?

Start here. Your home has an electrical service that delivers somewhere between 100 and 200 amps of current at 240 volts. That’s enough to run your fridge, your oven, your HVAC, and everything else, all at once. It’s roughly 24 to 48 kilowatts of capacity.

A single Nvidia Vera Rubin GPU, which starts shipping in volume in late 2026, is designed to draw about 1,800 watts. Not 1.8 kilowatts across your whole house. 1.8 kilowatts into one chip the size of a large postage stamp.

Now here’s the twist. The chip doesn’t run at 240 volts. Or 120. Or 12. It runs at about 0.85 volts. Less than a AA battery.

There's one equation you need, and then you're set for the rest of this piece: power equals voltage times current. That's it. That's the whole toolkit. If you keep the power fixed (1,800 watts) and shrink the voltage way down (to 0.85 volts), the current has to shoot up to compensate.

So how much current is that?

Roughly 2,000 amps arriving at a single chip.

That is ten times the current running through your entire house, delivered into a piece of silicon smaller than a saltine cracker. And a single AI rack has hundreds of these chips. A single data center hall has thousands.

That is the actual engineering problem. Everything else in AI hardware right now is a consequence of it.

Why is high current the enemy? The water hose analogy

Here’s an analogy that will get you the whole rest of the way through this.

Imagine you need to deliver a fixed amount of water to a fire. You have two choices. You can use a skinny hose at very high pressure, or you can use a giant fat hose at low pressure. Either one gets the water there. But the fat, low-pressure hose is enormously more expensive, because you have to build a much bigger hose, use much more material, and deal with much more friction and heat as the water sloshes through.

Electricity works the same way. Voltage is the pressure. Current is the flow. High voltage plus low current is easy to move around a building efficiently. Low voltage plus high current is a nightmare. It requires thick copper, it wastes energy as heat, and every foot of wire between the source and the destination bleeds off power.

And crucially: the losses in the wire go up with the square of the current. Double the current, quadruple the loss. This is not a small effect. This is why every high-power system ever built, from the national electrical grid to your car’s alternator, has always tried to move energy at the highest voltage it can manage and step it down at the last possible moment.

Now imagine trying to run a fire hose at zero pressure. That’s roughly what’s happening at the AI chip. The chip demands 2,000 amps at less than one volt. You have to get those 2,000 amps to arrive at exactly the right instant, at exactly the right voltage, without wasting the equivalent of a small hotel’s worth of energy along the way.

The way the industry is solving this is by redesigning every layer of the power path at once. Every layer. From the electrical grid outside the building all the way to the last inch of copper between a voltage regulator and the chip. That is what the AI capex cycle is actually funding.

The power path, from grid to chip, in plain English

Let me walk you through the ladder. Six layers. Skip the details if you like; the pattern is the same at every layer.

Layer 1: Grid to building. From AC to DC.

For twenty years, data centers took high-voltage AC power from the utility, stepped it down to 415 or 480 volts AC, and distributed AC around the building. Every server rack then had its own box that converted AC to DC. Every one of those conversions loses energy and generates heat, like every time you change money at an airport kiosk, you lose a little in the spread.

Nvidia has now announced, and multiple ecosystem partners are building around, a new architecture that takes utility power and immediately converts it once to 800 volts DC at the perimeter of the building. Then it distributes that 800-volt DC power straight into the racks. No more AC in the building. No more four or five conversion steps.

The results, per Nvidia and independent analysts:

• Up to four conversion steps eliminated

• End-to-end efficiency up more than 5%, which at gigawatt scale is dozens of megawatts recaptured

• Copper use down roughly 45%, because higher voltage means lower current means thinner wires

• Total infrastructure cost down roughly 30%

Nvidia has publicly partnered with a company called Navitas on the specialty power chips that make this possible. Vertiv is building the power modules. Foxconn is building an entire 40-megawatt data center in Taiwan around it. CoreWeave, Oracle, and Lambda are already designing to it. This is not a research paper. This is happening now, with money.

Analogy: Instead of every apartment in a skyscraper having its own currency exchange for foreign visitors, you put one exchange at the front door of the building and hand everyone US dollars from there.

Layer 2: The rack. From 40 kilowatts to a megawatt.

A rack is a metal cabinet about the size of a large refrigerator. It holds servers.

Five years ago, a rack full of general-purpose servers drew maybe 10 to 15 kilowatts of power, roughly the same as a couple of electric ovens running at once. An early AI rack running Nvidia’s H100 chips drew about 40 kilowatts. The current generation, Blackwell GB200, draws 120 to 130 kilowatts per rack. That’s the same as a small factory.

The next generation, Nvidia’s Rubin Ultra Kyber rack, which is targeted for 2027, is designed to draw 600 kilowatts to 1 megawatt. Per rack. That’s the electrical draw of about 500 homes, concentrated into a single cabinet.

Analogy: Imagine the electrical demand of an entire suburban cul-de-sac being served through a single garden hose. That’s the pressure the rack designers are dealing with.

Layer 3: Inside the rack. Moving power to each server.

Historically, the power inside a rack ran at 12 volts. A few years ago, the industry shifted to 54 volts because the current was becoming unmanageable. The next step, being designed now, moves the internal rack voltage to 400 volts, and eventually 800 volts DC. Same voltage as the whole building.

Same reasoning as before. The higher you keep the voltage, the less current you have to push through the copper busbars, and the less energy you waste as heat.

Layer 4: The server card. Getting power to the chip.

This is where the engineering starts to look less like wiring and more like sculpture.

The voltage regulators, the little boxes that do the final conversion from 12 volts down to 0.85 volts for the chip, used to sit next to the chip on the same circuit board. Current flowed sideways across the board to reach the chip. That worked fine when the currents were 50 or 100 amps.

At 2,000 amps, it doesn’t work at all. The board can’t carry that much sideways current without wasting huge amounts of energy in the copper.

So the industry is moving to something called vertical power delivery. The voltage regulators move to the back of the circuit board, sitting directly underneath the chip, and inject current straight up through the board into the silicon. The path shrinks from inches to millimeters.

The companies doing this work were sleepy specialty names five years ago and are central to the AI story now: Infineon, which showed its latest generation at a conference in early 2025, along with Empower Semiconductor, Vicor, and Monolithic Power Systems.

Analogy: instead of a food-court kitchen wheeling meals to tables on a cart across the room, every table gets a dumbwaiter that lifts the food straight up from a kitchen directly below.

Layer 5: The chip package. Reservoirs of energy.

Even with perfect power delivery, there’s still a problem. AI chips don’t draw power steadily. They gulp it. The load can swing from near-zero to 2,000 amps in about a millionth of a second. If the voltage regulator can’t keep up, the chip crashes.

To handle that, the industry uses capacitors: essentially tiny electrical shock absorbers, or, better, tiny batteries that can release energy in microseconds. Every AI chip is surrounded by thousands of them. The numbers are worth sitting with.

• A Vera Rubin rack contains about 600,000 capacitors. Per rack.

• AMD’s competing MI450 chip went through a late design change in which a single small capacitor part number jumped from 1,440 pieces per board to 10,544 pieces per board. That’s a 632% increase, because the design needed more shock absorption.

• Panasonic’s own investor materials from late 2024 disclosed that AI servers use 22 to 30 times more capacitors than general-purpose servers.

The industry is now moving capacitors off the circuit board and onto the chip package itself, to get them closer to the die. Samsung’s electronics arm announced a $993 million contract with an undisclosed US hyperscaler for specialty on-package silicon capacitors, with deliveries scheduled from 2027 through 2028. That is a real, boring, sub-industry most investors have never heard of, and it is about to be very lucrative.

Analogy: think of a Costco loading dock. If you’re moving pallets steadily all day, one forklift is fine. But if a truck shows up unannounced and has to be unloaded in ninety seconds, you need twenty forklifts standing by right at the dock. Capacitors are the standby forklifts, and AI chips need thousands of them within arm’s reach.

Layer 6: The chip itself. Regulators moving onto the silicon.

The endgame is to eliminate the last inch of copper entirely by building the voltage regulator directly into the same piece of silicon as the compute chip. Intel has demonstrated this on production nodes. TSMC has published a roadmap for combining voltage regulation and capacitors directly into the chip package in the 2028 to 2030 window.

If that lands as planned, the whole power-delivery problem gets absorbed into the chip itself, and the industry moves on to the next constraint, which is probably cooling.

Training vs. inference: why AI's economics are shifting to cost per answer

This is where the physics turns into an economic story worth caring about. AI hardware does two very different things. It trains models, the big, expensive, one-time process of teaching a system by grinding through mountains of data. And it runs them, which the industry calls inference, the part where you ask a chatbot a question and get an answer back. Training is what got AI to where it is. Inference is where the industry is going to make its money.

These two workloads have very different power profiles. Training is steady: you run flat-out for weeks, and the metric that matters is how much computing you get per watt of electricity, roughly like miles per gallon on a cross-country drive. Inference is spiky: millions of users asking questions at unpredictable moments, each request short, latency everything. The metric that matters there is cost per answer, or in the industry’s language, watts per token (a token is roughly a word).

For the last few years the whole industry optimized for training. That is changing fast. Nvidia’s language around the new Rubin chip has quietly shifted toward inference economics, with claims of ten times the throughput per megawatt versus Blackwell on inference workloads. Ten times is not incremental. It is a step change in what it costs to run a large model in production.

And that shift is exactly what drives the power-delivery arms race. Inference wants many smaller chips, running at lower voltages, packed closer together, drawing more current per square inch of silicon. That is precisely the recipe behind the 2,000-amps-at-0.85-volts problem this piece opened with. The economic pressure and the physical pressure point in the same direction, and both push hard.

Who wins the AI power buildout beyond Nvidia?

This is where the curiously-optimistic framing earns its keep, so bear with me while I get slightly evangelical. The innovation in this cycle is genuinely spread across dozens of companies working on completely different pieces of the same problem. This is not an Nvidia monoculture. It is closer to a symphony that Nvidia happens to be conducting. Here is my working map of who is doing what.

Nvidia designs the whole system and sets the reference; it tells everyone else what the target looks like. AMD is a genuine number-two on the chip itself and is pushing the same physics from a different direction. Broadcom, Marvell, and Alchip design custom AI chips for the hyperscalers (Google, Meta, Amazon, Microsoft), and every one of those chips has to solve the same power problem. TSMC does the advanced packaging that makes on-chip capacitors and, eventually, integrated regulators possible; it is the quiet linchpin of the whole story.

One layer out, Infineon, Empower Semiconductor, Vicor, and Monolithic Power build the power-delivery modules that sit inches from the chip. Navitas, Wolfspeed, Transphorm, and EPC build a newer class of power semiconductor made of gallium nitride and silicon carbide, which switches faster and handles higher voltage than traditional silicon; that is the enabling chemistry for the 800-volt building. Panasonic, Murata, TDK, Samsung Electro-Mechanics, and KEMET/YAGEO build the capacitors, all 600,000 of them per rack: boring, high-volume, and now a real dollar market. Vertiv, Eaton, Schneider, and the specialty HVDC-gear makers build the power infrastructure of the buildings themselves.

And the hyperscalers, meaning CoreWeave, Lambda, Oracle Cloud, and Together AI, plus Google, Meta, Amazon, and Microsoft, are now inside the reference-design conversation as co-architects. That is genuinely new. Data-center operators used to just buy servers. Now they help design the racks and the buildings.

Dozens of teams, different chemistries, different physics, different markets, all pointing at the same problem. That kind of setup, historically, tends to solve the problem, not on any one team’s schedule, but on a collective schedule that is usually short by historical standards.

What I take away as an investor and operator

Three things. First, the binding constraint on AI infrastructure for the next five to seven years is power delivery, not transistors. Chipmakers are still shrinking transistors, but per-chip power draw is climbing faster than transistor efficiency is improving, and that gap gets closed at the package, the card, the rack, and the building. If you want to know who wins the AI capex cycle after Nvidia, look at the companies whose products live on that closure path. Many of them are unglamorous, most are not household names, and some are worth serious money.

Second, the shift from training-first to inference-first economics changes which architectural moves matter and which financial metrics to track. Watts per token and cost per answer are the numbers that will dominate AI earnings calls three years from now. Anything that improves the efficiency of running a model, as opposed to training one, will get rewarded: better power delivery, but also better memory architectures, better cache management, better decoding. If you are evaluating an AI-infrastructure investment, look at its inference story, not just its training benchmarks.

Third, and most important, this is fundamentally an optimistic story. The industry is not stuck. The bottleneck is well understood, the physics is not exotic, the roadmap is credible, and it is being funded with real money right now by the biggest and most sophisticated companies in the world. When that many well-capitalized teams converge on a physical problem with an obvious return on investment, the problem tends to get solved. Not perfectly, and not on time, but faster than almost anyone outside the industry expects.

If the last decade of AI was about proving the algorithms scale, the next decade is about proving the electricity does. That is a far more tractable problem than it looks, and it is where I would spend my attention.

Curiously Optimistic is a newsletter about specialty chemicals, advanced materials, industrial operations, and AI, written from the operator’s line. If a friend or colleague forwarded this to you, you can subscribe at the top of the page. If you want to talk about anything in this piece, my calendar is open.

Bob Girton is the principal at August Grace Advisory, where he leads investment and strategic advisory work in chemicals, advanced materials, and AI infrastructure.

Sources and further reading

• NVIDIA 800V HVDC Architecture Will Power the Next Generation of AI Factories, NVIDIA Technical Blog

• NVIDIA 800V HVDC Deep Dive: From 54V Racks to Megawatt AI Factories, SanYi

• NVIDIA, Partners Drive Next-Gen Efficient Gigawatt AI Factories, NVIDIA Blog

• NVIDIA Rubin Platform Begins H2 2026 Ramp, Let’s Data Science

• NVIDIA Selects Navitas for 800V HVDC Architecture, Yahoo Finance / GlobeNewswire

• NVIDIA AI Server Power Roadmap: Kyber’s Next Generation Strategy, Ming-Chi Kuo / Medium

• The New Power Stack for AI Servers, SemiVision

• Conductive Polymer Capacitor Market and Design-In Guide to 2035, passive-components.eu

• NVIDIA Rubin Platform, Open Models, Autonomous Driving, NVIDIA Blog