That much is settled and it is now the consensus, with Lex, Nocera, Novogratz and Rubinstein all landing there in the past week. A levered book is marked every night by somebody else, and when the mark goes far enough against you the decision to sell stops being yours. Chamath Palihapitiya called it "sort of an automatic one-way ratchet" on All-In.

None of it tells you whether the analysis was any good.

Subscribe now

He published the thesis before he raised against it. Situational Awareness, June 2024, argued that AI scaling would run into physical limits before it ran into intellectual ones. The fund took the essay’s name and its positions, and ran to about $45bn.

His thesis had two legs and they need grading separately.

↑ Long the physical build-out of AI.

↓ Short the application software AI was supposed to eat.

One can be graded against disclosure today. The other cannot be settled until the capacity is delivered and priced, however well the demand marks in the meantime.

The build-out leg

This one marks well, and better than I expected.

In June 2024 the essay put the largest 2026 training cluster at about 1 GW and about 1 million H100-equivalents, and Epoch AI's tracker now has Colossus 2 at 946 MW and 1,112,000, inside Epoch's own confidence interval on power and about 11% out on chips. Epoch measures campuses. The essay forecast a training run.

On 2026 AI capex it said near $500bn against a big-five outturn closer to $800bn.

He was low.

Underneath sat a call that power would bind before chips did, contrarian in mid-2024 when the supply conversation was Nvidia allocation and advanced packaging. It reads as the consensus framing now.

GE Vernova held 116 GW of gas equipment committed at 30 June, against 100 GW a quarter earlier and 83 GW at the end of 2025, and it is booking deliveries into 2031, expects to be more than halfway contracted for that year by December, and has 2032 discussions underway. That is a demand signal you can bank.

Read what it is made of. 53 GW of the 116 is firm orders. Slot reservation agreements are the other 63 GW, contracts for a place in a queue, and they have run ahead of firm orders in every quarter of that series. Queue positions are 54% of the book.

Power is also where his chapter went wrong, one layer down. He treated power as a generation and fuel problem soluble with shale gas and turbines, and GE Vernova's own chief executive says otherwise. On the first-quarter call, reported on 22 April, Scott Strazik said gas turbines are "really not the gating item" in many of these projects.

Strazik's reason: "a 3-year cycle from when a project starts, the EPC buildout, the permitting, the fuel availability". Three years is the project clock. A turbine is one line inside it.

Widen it past Strazik's three and the constraints multiply. CBRE puts interconnection queues at 24, 36, 48 months and beyond, and reads the queue as binding. PJM's 2028/29 capacity auction cleared at its administrative cap of $325 per MW-day and still came in 6,831 MW below the reliability requirement. AEP Ohio quotes seven to ten years to fully catch up with load at Meta's Prometheus campus. Curtailment rules for large loads arriving without their own supply are the fourth.

Those clocks start in different places and cannot be ranked against each other. In the operating record, power is a queue problem, a market-design problem, a transmission problem and a curtailment problem. His chapter addresses generation and fuel, and none of the four.

The software leg

This one lost, and it lost badly.

has done the arithmetic: the 29 disclosed US longs fell 21% on average and shorts such as Adobe rose 26%. Porter Stansberry has gone further and argued the software thesis was simply wrong. Both legs went against him at once, and that is before the leverage.

What the tape cannot tell you is whether the short's mechanism ever existed, and it needed a specific one: AI does the work, the customer needs fewer licences, the vendor's revenue falls. On the disclosed record it is running backwards. Microsoft reported over 30 million paid Microsoft 365 Copilot seats on 29 July, against his own 2024 illustration of a third of roughly 350 million Office subscribers paying $100 a month for an add-on. Workday's contracted user base went from more than 70 million in January 2025 to more than 80 million by May 2026.

On the tracked set, evidence against that mechanism is broader than evidence for it. Of 24 companies tracked through this argument, 17 publish a retention metric and 11 of those are flat or rising over the last four quarters. Altimeter's public-SaaS median went from 108% to 110% over the reported quarters, and Meritech independently had the sector at 108% in March.

Two more readings come from inside the companies they describe. Atlassian sells the seats it is reporting on, and last October it studied its own base: users of AI coding tools expand their paid Jira seats about 5% faster than non-users. Figma reported on 5 August in the category the thesis said would go first, revenue up 48%, net dollar retention 136%, down from 139% the quarter before, and by its own count a third consecutive quarter of acceleration, its own headline crediting sustained seat expansion and AI adoption.

Figma wrote that about Figma.

Against that, GitLab has fallen from 130% net retention to 117% across nine consecutive quarters without a reversal. ZoomInfo and HubSpot, the two closest to pure per-seat licence capture, both broke in the same window. ZoomInfo printed 89% retention on the same day and took the first goodwill impairment in its history.

One measurement comes closest to sorting it. Aleph and Benchmarkit's joint benchmark of 1 June 2026 covers 342 B2B SaaS and AI-native companies, 230 of them reporting retention, on full-year 2025 actuals. Split by pricing model, median net retention runs 98% seat-based against 108% usage-based, with an all-company median of 102% across those same 230. Their stated reason is the short's own mechanism. Seat-based pricing faces the opposite pressure, they write, because in an era of AI-driven headcount efficiency, customers consolidating roles or deploying AI agents shrink their own seat count.

Three things bound that. Aleph sells AI-native planning software, so the cause it reads out is the cause that sells its product. A single cross-section cannot separate AI-driven seat compression from budget scrutiny, cohort maturity or an infrastructure-against-application split. And the instrument is a year old.

The seat is under pressure. Retention across those 24 is holding. Most of the vendors winning on usage pricing are infrastructure, close to the boundary he drew himself. Figma sells seats and credits together, so it bears on where the application boundary sits and does not settle the pricing question.

He drew his line in roughly the right region and on the wrong axis.

Seat against meter is what moves this data. Application against infrastructure is what he shorted.

The 6 August prints cut against that in both directions. Atlassian's cloud revenue accelerated to 31%, and it named strong seat expansion in core Jira and Confluence as the first reason. It then guided the coming year's cloud growth to 25.5% against the 27.9% it had just delivered. Datadog, priced on consumption, raised its guidance and closed down 19%, having disclosed that its largest customer, an AI-native, cut usage from the third quarter of 2026, the quarter now running.

I have not established that the disclosure caused the fall, and one session is not a pattern. What it shows is that the axis I have just drawn is thinner than the preceding week made it look, and that an AI company cutting its software bill lands on the meter, which is the side I had marked safe.

What a demand forecast settles

Neither verdict touches the question everyone is actually asking, which is whether the build-out leg pays. That leg has no instrument. Two earlier physical build-outs are the closest thing to one, and they bear on it in a way that has nothing to do with being wrong.

Between December 1996 and December 2002, US internet backbone traffic grew roughly 73 times, and real consumer spending on telecoms services rose 72% between 1995 and 2001. From its March 2000 peak to 16 May 2003 the Nasdaq Telecommunications Index fell 89%. By September 2002, 23 telecoms companies had gone bankrupt and roughly $2trn of market value had disappeared. Odlyzko's verdict is one sentence: the telecom crash was not caused by a decline in traffic growth.

Britain's railways did it first. Route miles in service rose 171% between 1845 and 1850, total revenue 108%, all on Odlyzko's annual series, and by 1860 revenue had more than quadrupled from 1845. Equity peaked in August 1845 and fell 67% to its trough in April 1850.

In both, capacity outran what it could earn. Fibre was overbuilt and repriced, wholesale bandwidth prices already falling 20% to 30% a year in 1999 and early 2000 and 40% to 60% on competitive routes, all of that before the peak. Britain's railways were overbuilt against business plans Odlyzko names as over-optimistic and as the cause, at £31,700 to £35,700 a mile against a £20,000 promise, which Odlyzko puts at 50% to 100% above plan across the industry.

Financing turned that into ruin, and it is the part that belongs in a piece about leverage. Railway shares were sold partly paid, so a subscriber put down a deposit, took the share, and stayed liable for calls on the rest as construction proceeded. Median paid-up capital on a new railway stood at 6.0% at the end of 1845, 13.1% a year later, 39.1% by 1847 and 67.2% by 1848, so at the outset a subscriber had committed to roughly 16 times what he had handed over. Investors were called for capital through the years the equity fell, and the calls arrived with the break and kept arriving for years after it.

I am not drawing a rule from two cases, and they were chosen as the closest physical-build analogues available. They differ from this one in the place a reader will reach for first. A railway subscriber was liable for calls on demand, and the fully-paid shares that fund this build-out carry no such obligation. What crosses over is the capacity, not the capital structure, and the capacity is the half a demand forecast speaks to.

The claim the two cases support is narrower than that.

Volume being right is compatible with equity being destroyed, and nothing in a demand forecast addresses either price or the terms on which the build is financed.

GE Vernova's reservation book is, it seems to me, the live version of that question. Slots contracted against delivery constraints that carry no date.

What would settle it

His software leg needed a mechanism the disclosed record does not show, and the instrument that would settle it is thin. A weak instrument bounds how hard a verdict can be pushed and says nothing about its direction. Of the 24 tracked, three publish no retention measure of any kind and four publish something else. Adobe, the short confirmed by name in the first wave of reporting, discloses no retention figure in its most recent filings.

His build-out leg is a demand and capacity forecast, marking at or ahead of his line, saying nothing about what a gigawatt of contracted capacity earns once delivered. Its clock runs in interconnection queues and transmission build, which nobody has dated and which runs a great deal longer than the financing he had. He named the right category, addressed the one link his chapter is about, and never asked the question the two cases keep putting: whether the volume converts into returns.

He has now put money somewhere, and the choice is more specific than a retreat from leverage. Situational Awareness put a reported $400m into Source Foundry on 4 August, after a reported $100m in July, taking the position to about $500m on the figures Bloomberg and the Wall Street Journal carry. None of it appears in a filing. The firm has told investors it will no longer borrow from banks to magnify bets.

Source Foundry registered in California on 18 July 2025, incorporated in Delaware, and lists an address in San Francisco. I could find no company website, product page or statement of its own.

What it is building comes from its backer. Sequoia's Stephanie Zhan says Source Foundry is targeting a bottleneck in semiconductor manufacturing by developing tools for lithography, and that it means to use different underlying physics to raise the speed and scale of chip production. She calls the mismatch the chip wall. She never says ASML and she never says EUV. That ASML framing, which ran through the week's coverage, is the reporters' addition, laid on a template a different and better-documented startup had already built.

Read that against the book he was carrying in the spring.

His first-quarter filing was 62% put options, and among its eleven put lines sat $494m of notional against ASML, matched by $6.1m of ASML common. ASML was the only semiconductor-equipment name in the whole book. Whether that put was still on when the Source Foundry money went out is not on the public record. The two would express a single view only if Source Foundry turns out to substitute for what ASML sells, and neither the overlap nor the substitution is established.

His essay is a harder comparison than it first looks. It called power the binding constraint, and it named advanced packaging and high-bandwidth memory as the primary constraints on the chip side, so it did not wave past chips. What it does not mention once is lithography.

His next reported private investment targets it. Whether that reflects a revised view of where the chokepoint sits, a bet on a scarce supplier, or both, is unknown. An investment shows a chosen exposure without revealing how he now ranks power against lithography, and the two run on different clocks: the essay is about scaling to 2028, and a tools venture pays out in the 2030s or never.

Both tranches went somewhere with no daily public mark. No listing, no exchange print, no prime broker. A private position is still valued periodically for the fund's own books, and none of those valuations can force a sale overnight. After roughly 21 months in a book marked every night by somebody else, that is the whole of what he bought.

Three disclosures would move this. On the build-out leg, whether GE Vernova's reservations keep running ahead of firm orders, which tells you demand is real and nothing about price. On the software leg, a seat-against-usage retention split published on a stated period, since the one we have carries none. And on 14 August the second-quarter filing falls due, reporting positions as they stood on 30 June, which will show whether the ASML put was still on at quarter-end and nothing about what survived July. I suspect pricing-model disclosure settles the software half well before revenue does, and the build-out half not for years.

The leverage is over. Neither leg is. Nor is the fund.

So the usual measuring instruments are missing. Travis Kalanick supplies an alternative one himself.

His framework, published on atoms.co, treats the physical economy as a computer.

Manufacturing is the CPU, Real Estate is the storage, Transportation is the network.

He pushes it further on the a16z Show. A CloudKitchens building becomes "a 10,000 square foot semiconductor", a "30 core processor" with 30 kitchens. The corridors where the food moves are a network bus. The courier who brings the meal is "a TCP packet".

Take that as an architecture, and it is the only available test of what $1.7bn has bought, because Atoms wrote the specification.

Atoms' own site names six operating companies across three divisions. Five sit in Food, one sits in Mining, and the Transport slot is currently absent.

Five, one, zero, under a three-layer design presented as balanced.

Subscribe now

Key insights

• The metaphor is the instrument: Kalanick's own CPU, storage and network framing grades his three layers very differently.

• Storage is the part that exists: CloudKitchens and ProFood Properties are property businesses at working scale, on a mixed operating record.

• The network is one quarry and a placeholder: Pronto has hauled more than 2m tons autonomously. Atoms Transport has no company, product or customer.

• The CPU lives in job adverts: world models and vision-language-action models appear only in job postings for a founding team, and the engineering blog has never covered a model.

• Owning the assets has a rationale and a ceiling: co-design is his own argument for it, and industrial margins are what it earns.

What Atoms actually is

Atoms is City Storage Systems with a new name, and the sequence matters more than the rebrand.

On 13 March 2026 Kalanick revealed the brand and the holding company on TBPN, saying the business had spent eight years in stealth. In early April Atoms completed its acquisition of Pronto. On 22 July, in his words, the round "merged our various businesses into one Atoms equity structure".

Read March as a reveal. Nothing was bought that month.

CloudKitchens was founded in January 2016 as Urban Kitchens. Kalanick's 10100 fund bought control of the parent with about $150m in March 2018. Otter was built in-house in 2020 and Lab37 in 2023. Of the six named companies, Pronto is the only external purchase.

The capital already in the business is larger than the headline suggests. Saudi Arabia's Public Investment Fund put in $400m in January 2019. November 2021 brought about $850m of mixed debt and equity at a reported $15bn CloudKitchens valuation, split undisclosed. That $15bn belongs to a predecessor entity and it is not an Atoms valuation.

Tracxn puts pre-2026 funding at about $1.25bn and PitchBook at $1.75bn. The trackers do not agree, and the instruments do not compare. That is part of why so little about Atoms can be pinned down.

The opacity extends to the cap table. TechCrunch reported Uber's cheque at $100m, written around January 2026, so part of the July announcement is the disclosure of an earlier investment. Uber's 10-Q for the quarter to 31 March 2026 does not mention Atoms, CloudKitchens or Kalanick. The $100m sits unnamed inside a $179m line for non-marketable equity purchases, against $13.2bn of quarterly revenue. Every strategic reading of that cheque is inference.

Two governance facts belong on the record.

Kalanick resigned as Uber's chief executive on 20 June 2017, after the Benchmark-led "Moving Uber Forward" letter and the Holder review. Anthony Levandowski, who founded Pronto and is named by Chemistry as part of the Atoms founding team, pleaded guilty to trade-secret theft and was pardoned in January 2021.

Storage is the layer that exists

CloudKitchens converts marginal commercial property into blocks of small delivery-only kitchens and rents them to restaurant operators. Its own site claims 90-plus US facilities and 20-plus kitchens per property. Chick-fil-A, Taco Bell, Wingstop and Popeyes appear in the network, and Chick-fil-A confirmed its June 2026 Wynwood site as its sixth.

ProFood Properties reports more than 500 sites and $715m of assets under management, self-reported. The two counts should not be added; nothing establishes whether they are separate buildings. Otter sits on top as software, claiming 100,000-plus restaurant locations. Atoms has published no revenue for any unit. This is a real business with paying customers on two clocks, kitchen tenants and software subscribers.

The operating record underneath is mixed, and it carries dates.

The Financial Times reported CloudKitchens buildings around half full at the end of Q1 2023. Restaurant Business, reviewing 20 US locations, calculated tenant churn of 65% within a year in 2022. East Oakland closed in November 2025 and went up for sale. That evidence is two to four years old, and the estate has expanded since.

Ask Kalanick what protects the business and he answers property.

On TBPN he said a competitor would need "billions of dollars of real estate in every major city in the world".

He did not answer data. He did not answer models.

The network layer, one quarry and an empty slot

Pronto is the one part of Atoms with third-party-verified deployment.

Levandowski founded it in 2018 to build highway-truck driver assistance. It moved off-road because quarries carry lower capital, R&D and regulatory burden, and bought its rival SafeAI in July 2025.

At Heidelberg Materials' Lake Bridgeport quarry in Texas, a mixed fleet of Caterpillar 775G and Komatsu HD605 trucks hauled more than 2m tons of limestone autonomously in under eight months. The system is vision-only: cameras, satellite positioning and drive-by-wire, with no lidar, no radar and no pre-mapping. Komatsu sells it through its dealers as Komatsu Smart Quarry Autonomous. Mariana Minerals has run it at Copper One in Utah since April 2026.

That is industrial autonomy earning on named sites. Heidelberg published the tonnage itself, and no other Atoms operation has an output figure disclosed by its customer.

Now scale it honestly.

Heidelberg's plan is 30 autonomous vehicles by the end of 2026 and more than 100 by the end of 2028, a dated ramp target. Pronto is recruiting safety drivers to sit in the trucks, and the unmanned share of the fleet is unstated. The Hitachi agreement of 16 July 2026 is a memorandum of understanding, with the first joint project not yet running.

Kalanick describes the sales call himself. Offer a gold mine chief executive 20% more gold a year and nobody says no… "but they say prove it."

Then there is the division named after the layer. Atoms Transport has no company, no product, no customer and no site. Its slot reads "Announcement soon", with placeholder text lifted from Lab37 days after a $1.7bn raise. A third of the stated architecture is a page that has not been written.

The CPU is a hiring plan

Kalanick concedes the gap, and he concedes it in his own signed manifesto.

The atoms.co vision page states that "this kind of intelligence requires computation we haven't invented, at an efficiency we can't yet fathom". The framework predates Atoms by a year. At the Abundance Summit in March 2025 he said "atoms AI" needs "a whole other set of models that are going to need to be invented". The thesis has not moved since.

Follow the concession into the record and it holds. World models, vision-language-action models and reinforcement learning appear in exactly one place in Atoms' public footprint: job specifications. Both senior postings describe the hire as joining "our founding team", in July 2026. There is no published model, no paper, no benchmark and no custom silicon, and the deployed stack is off-the-shelf NVIDIA Jetson.

What does exist is engineering. Fleet-log pipelines over multi-petabyte datasets. A labelling operation run by an analyst managing remote labellers. The deployed mining stack is classical control, Kalman filters and Pure Pursuit path following, with learned perception on top.

The published output points the same way. Atoms runs an engineering blog whose About page names "frontier physical AI models" as a focus. Its eight posts since September 2025 are all backend distributed systems, and so are its seven public repositories.

Not one line on robotics, perception or a model.

A company shows you what it has built by what it can afford to publish.

The senior bench has the same shape. Brian Attwell, the group CTO, is a distributed-systems engineer from CloudKitchens. Eric Meyhofer, Kalanick's "head of the Advanced Technology Group at Uber", runs Lab37, and Pronto's Catherine Culkin is the most relevant autonomy hire. No head of AI and no chief scientist is named at group level.

The showpiece of the CPU layer is a dispensing machine, and Kalanick calls the category "gainfully employed robots". Lab37's Bowl Builder produces up to 300 bowls an hour on its 2026 specification, against the 80 Restaurant Business reported in December 2023.

Humans still prepare and load every ingredient. On a recent podcast, he put labour at about 30% of revenue in the delivery kitchens, and "between 7% and 10% of revenue" when the machine runs.

The customer list is the harder fact. None has ever been named, there is no unit count and no price, and the site still invites deposits for early-production slots. The only verifiable installations sit inside Atoms' own virtual brands, in Pittsburgh, Los Angeles and New York.

A machine with no named outside customer is a pilot.

Three engineering programmes sit under one phrase. "Atoms' physical AI" names classical off-road autonomy in production, a state-machine dispenser at one site, and a world-models programme that exists in job adverts.

One company, three engineering cultures, and no shared stack in evidence.

What owning the assets costs

The strongest case for owning is his own, and it has nothing to do with data.

Kalanick's argument is physical co-design.

In a 2025 podcast, he put it this way: "these kitchens are designed for these kinds of machines to be in them and vice versa." An incumbent cannot do that.

Replacing the front line in a Chipotle means demolishing it, and then "my store's down for two to three months, and the economics start to not work". That is a real advantage, and only the greenfield operator gets it.

The rest of the physical-AI top tier sells into other people's operations. Prometheus raised $12bn in June 2026 at about $41bn, and Figure carries $39bn post-money. Labs and robot makers.

Atoms buys the operation.

What the choice does for certain is fix the revenue line as operating revenue in industries with industrial margins. CSIMarket put transport and logistics operating margin at 7% in Q1 2026, and freight forwarding net margins at 3% to 4%. Heidelberg Materials, Pronto's anchor customer, reported first-quarter 2026 operating EBITDA down 13%. A per-site payback figure would show otherwise, and Atoms has published none.

The failure mode also changes shape. A vendor that fails to sell loses a sales cycle. An owner that fails to fill loses rent, depreciation and interest. CloudKitchens at half occupancy in early 2023 is the demonstration inside the group's own history.

Katerra took about $2bn from SoftBank into vertically integrated construction and went bankrupt in June 2021. Zume raised $445m for robotic pizza and liquidated in 2023. Convoy peaked at $3.8bn and shut in a 2023 freight downturn. Katerra and Zume are the reads on Atoms Food, Convoy on a leveraged asset owner meeting a cycle.

a16z's own 2024 essay on AI for the physical world argues for selling AI into complex, safety-critical hardware environments. Atoms buys and runs the environment.

What would change the reading

This method is portable. When a company hands you its own architecture, audit it layer by layer and count which layers are staffed.

→ Its storage layer collects rent across two property estates.

→ Its network layer runs a quarry in Texas and a copper mine in Utah.

→ Its CPU layer has 200-plus open roles and a founding team that has not started.

Four checkpoints would settle most of this.

A named third-party Bowl Builder customer, with a unit count and a price. The unmanned share of the Heidelberg fleet. The size of the bank facilities. The company on the Transport slot.

There is one more move to watch, and it runs the metaphor backwards.

Kalanick said in June 2026 (All-In Podcast) that his 500 properties already carry power, cooling and piped gas. He is "literally looking at putting some of our compute into our kitchens", and on power-dense properties, "those are like hope diamonds ... which I own".

The Storage layer is reaching for the CPU layer, with the one asset he definitely has.

Subscribe now

On Friday 10 July, from Wolfsburg, Volkswagen set out its strategy to 2030.

It will halve its model line-up, cut variants and powertrain options by up to 75%, and take annual capacity from about 10 million vehicles down to about 9 million. On the number everyone was waiting for, the jobs and the plants, it said nothing.

That silence is the story.

The day before, the supervisory board committee voted 12 to 7 against management's restructuring package. That package held the job cuts, the German plant closures, and a spin-off of the core VW brand. Labour and Lower Saxony carried the vote.

So the Friday announcement held only what management can execute on its own. Product line-up, variant count and capacity need no board consent. Plant closures need a two-thirds board majority, and the board had just refused. Blume spoke of shedding "excess capacity" as the world "deteriorated over the past 12 months." He named the disease and left the treatment unspoken, because the treatment is the board's to grant.

That makes this an industrial-strategy story, and it turns on one institution.

German co-determination seats labour on half of VW's board, and the 1960 Volkswagen Law hands Lower Saxony, on about 20% of the votes, a legal blocking minority. The institutions that protect German industrial jobs also set the maximum speed at which German industry can adjust.

So VW restructures at the pace of its own negotiations. The market it competes in moves faster.

My read: The governor just became visible.

Everyone knew European volume car-making was in structural trouble; last week we watched the mechanism that decides how fast it is allowed to shrink. The measures VW did announce still bite, so the workforce erodes through model death and attrition while the headline number stays unsayable. Management's real lever now is the corporate-structure workaround: spinning the brand into an entity outside the Volkswagen Law's fence, which turns a jobs fight into a constitutional one. The 100,000, still only reported and never confirmed by VW, is the least of it.

Toyota and Micron added US jobs and capacity behind the tariff wall the same week; Europe's champion cannot even announce its own retreat.

Forward Prediction: VW's next fight is constitutional. Management pivots from the headcount to a legal push to move the workforce outside the Volkswagen Law, and that becomes the battle of the autumn.

The 10 July announcement held only the board-independent measures. The jobs and plants the board vetoed the day before stayed unspoken.

Euronews: Volkswagen unveils four-year plan as questions remain over jobs and plants

On 8 July Blue Origin confirmed it is raising about $10bn at a $130bn valuation, the first external capital in its 26-year life.

Coatue leads with roughly $4bn. Bezos, who has funded the company by selling around $1bn of Amazon stock a year, adds about $2bn. The raise lands weeks after New Glenn exploded in testing in late May, destroying Launch Complex 36. It comes a month after SpaceX's IPO cleared more than $85bn at a $1.75tn valuation. The cash rebuilds the pad and funds a stated 50 launches a year by 2030, seven Blue Moon landers, and the TeraWave broadband constellation.

For 26 years Bezos's chequebook removed capital discipline as a governor. Outside equity puts it back.

Coatue's money arrives with a valuation, a cap table, and in time a clock. Two things follow. Blue Origin is now priced at roughly 7% of SpaceX, so every investor conversation becomes a relative-value trade. The pressure on Dave Limp is to show New Glenn cadence soon. And the constellation and the lunar cargo both depend on a heavy-lift vehicle that just blew up. The $10bn buys runway to prove reusability before the manifest slips and customers walk to Falcon or Vulcan.

The raise is as much about New Glenn's credibility as about the balance sheet.

CNBC: Blue Origin raising $10 billion at a $130 billion valuation, its first outside capital

Three deals in one week point at the same patch of ocean floor.

Thales agreed to take 35.5% of French maritime-drone maker Exail at €134 a share, valuing it near €3.9bn, with full control by early 2028. Italy's Fincantieri agreed to buy four underwater specialists, reported at around €600m. In London, Kraken Technology Group closed a $175m Series B at a $1bn valuation, Europe's newest maritime-defence unicorn, backed by the British Business Bank, the NATO Innovation Fund and Rheinmetall. And Canada's Kraken Robotics agreed to buy Britain's Covelya group, owner of Sonardyne, for C$615m.

The binding constraint is the seabed itself, now contested infrastructure.

After Nord Stream and repeated Baltic cable cuts, Europe cannot monitor or defend the cables and pipelines it runs on. Each of these deals buys a piece of the sensing and autonomy stack to close that gap. Two things stand out. The buyers are underwriting the bet with dual-use civil cash flow, offshore survey and cable inspection. That civil revenue is what makes Exail's €3.9bn tag defensible on defence budgets that have not yet arrived. And the consolidation is happening ahead of the demand wave, so the field of independent European undersea specialists is being priced and absorbed now.

British subsea IP passing to a Canadian owner is a sovereignty question the UK has so far waved through.

Breaking Defense: Thales acquires Exail Technologies stake as it expands undersea warfare footprint

On 6 July Toyota said it would spend $3.6bn on a second assembly line at its San Antonio plant, adding 2,000 jobs.

Tacoma production moves out of Baja California in Mexico over about four years. Cumulative investment at the site reaches $8.3bn. Bloomberg read it plainly as cutting tariff exposure on trucks.

Baja California is a capable plant, so reshoring the Tacoma is a more expensive way to build the same vehicle. Toyota is paying a premium to sit inside the US tariff line.

Two consequences follow. The line moves from Mexico to Texas and no new capacity is created, so 2,000 Texas jobs come substantially at Mexico's expense. That is a visible marker in the USMCA renegotiation, and a template other carmakers will be pressed to copy. And the four-year transition keeps Baja running throughout, which buys Toyota the option to slow or reverse the move if tariff policy softens.

Read it against Micron's wafer deal later in this edition, and the tariff regime is now a line item in every industrial capital plan.

Toyota: Toyota announces $3.6 billion expansion and 2,000 new jobs at its San Antonio plant

The China Construction Machinery Association released first-half data on 6 July.

Excavator sales ran to 152,320 units, up 26.4% on the year. Exports grew 33.5% to 73,295 units, roughly half the total, and June alone was up 35.3%. Loaders followed, with exports up 39.5%. Electrification is still marginal, 135 electric excavators in the half's exports, and it is visible in the numbers for the first time.

China's domestic construction market has been flat for years, so the entire growth story is exports.

That leaves the sector acutely exposed to the tariff walls and local-content rules in its destinations, US Section 301, EU trade defence, emerging-market content quotas. That plays out in two ways. XCMG, SANY and Zoomlion are being pushed to build inside their end markets, the same force reshoring Western machinery, running in reverse for Chinese exporters. And the export surge is a margin-mix story as much as a volume one: Chinese makers are climbing into electric and autonomous ultra-class equipment, the segment Caterpillar, Komatsu and Liebherr have defended on price.

The 135 electric export units are tiny, and they are the leading edge of a deliberate move up the value curve.

H1 2026 excavator sales, split domestic and export. The domestic market is flat; the growth is overseas.

SCMP: China's excavator exports surge 33% in first half as industry pivots to electric power

Two markers frame the week.

XPeng is pushing its IRON humanoid to mass production by end-2026, with founder He Xiaopeng running the unit personally. IRON shares the VLA 2.0 AI foundation with XPeng's cars. It will be built on the supply chain that ships more than 190,000 EVs a year, in a new 110,000 m² plant. Days earlier, on 3 July, Unitree won approval for a roughly $619m Shanghai STAR Market listing, on 2025 revenue near $250m and 5,500 humanoids shipped. In the US, Agility Robotics agreed to go public through a $2.5bn SPAC with a ~$200m Foxconn-led PIPE.

The binding constraint on humanoids is manufacturing scale and unit cost, and China is dissolving it through the car industry.

Automotive-grade parts, domestic AI silicon and existing high-volume lines let XPeng and Unitree price and build at a cadence Western pilots cannot match. Two things follow. Public-market capital and carmaker balance sheets are putting Chinese humanoids onto real floors. Agility's SPAC-plus-strategic structure is a tell that private markets would not fund its scale-up at that price. And the shared foundation across robot, car and flying car is platform economics, spreading the model cost across three product lines.

The mass-production headlines are still showroom trials; commercial scale is a 2027 test.

Caixin Global: Unitree Robotics wins approval for $618 million STAR Market IPO

On 6 July San Francisco's Katalyze AI raised a $10.5m seed.

Its agentic operating system stitches the silos of pharma manufacturing, the MES, LIMS, historians and SAP, into one operational record. Agents then run engineering and manufacturing tasks on top, with full traceability. It names Sanofi and says five of the world's 20 largest drugmakers use it. It cites an analysis that normally takes a year and $4-6m done in 45 minutes.

In regulated manufacturing an agent earns its place only if every action it takes is auditable for the FDA or EMA.

So the moat is the validated, compliant data layer across legacy systems, and whoever owns that layer owns the agents above it. That has two consequences. Landing five of the top 20 pharma names before scale is the real signal, because pharma is the hardest agentic beachhead precisely because of the regulation. And the same contest is opening in general industry: Britain's HIVE raised $15m the next day for a horizontal "silicon brain" to run any machine. Oxfordshire's Luffy AI took €9.4m for a control layer that also cuts the energy a motor draws.

The models are commoditising; the trusted operational record beneath them is the prize.

Tech Startups: Katalyze AI raises $10.5M to bring agentic AI to pharma manufacturing

On 9 July Micron lifted its planned US investment to more than $250bn through 2035, up from $200bn and $170bn before that.

It poured first concrete at its Clay, New York megafab a quarter ahead of schedule. The number that carries the weight sits lower in the release: a $500m financing and 10-year supply deal with GlobalWafers for 300mm raw silicon made in Sherman, Texas.

Almost all AI-memory coverage counts fab dollars and HBM bit growth. The quiet move is Micron underwriting a decade of domestic polished-wafer supply.

HBM stacks use far more wafer area per usable bit than commodity memory. The AI memory cycle is silently a demand shock for the substrate the whole industry starts from. Two signals sit under that. By financing GlobalWafers' US ramp, Micron is de-risking the one input a fab cannot import quickly under a tightening trade regime. Raw silicon sits even more concentrated in East Asia than the fabs it feeds. And the schedule acceleration is its own signal: pulling a megafab forward is rare and costly. It says Micron is chasing a demand curve it does not want to be short on.

The $250bn is the headline; the $500m is what holds it up.

Micron's escalating US commitment. The load-bearing line is the $500m raw-wafer deal at the base of the stack.

DatacenterDynamics: Micron increases total US chip investment to $250bn, announces $500m GlobalWafers partnership

A critical driver we’ve talked about before, building power for AI.

Alberta got a $3.2bn gas plant, National Grid Ventures put $1.75bn into a Microsoft campus, and Washington lent $3.26bn for AEP Texas transmission against a 41 GW load queue. A Duke University analysis says most of that power already exists. If large loads accept curtailment for 0.25% of the year, about two hours at peak, the existing US grid absorbs close to 100 GW of new demand with nothing built. The headroom runs to roughly 300 GW.

The wall in front of a data centre is the interconnection queue; the megawatts are already there.

GridCARE scans a utility's network with AI to find the slack and pairs it with flexible tariffs. It promises a connection in 3 to 5 years against a decade in the queue. It cites 650 MW freed on National Grid's New York system and 400 MW at Portland General Electric. Two things follow. The gas plants are the expensive answer to a problem that is half administrative, priced against capacity already built. And the real constraint is trust: whoever gets hyperscalers comfortable being interrupted, and regulators comfortable they will be, frees more capacity than any turbine order.

The cheapest capacity last week was the capacity already on the grid.

The week's new-build announcements against Duke's ~100 GW reachable at 0.25% curtailment. The cheapest capacity is the capacity already built.

pv magazine: Software can unlock 300 GW of capacity on the US grid without building a single power plant

Two governments moved on the input cost of reindustrialisation last week.

Britain's Modern Industrial Strategy turned one on 8 July, and its sharpest line raised the discount on electricity network charges for energy-intensive industry, steel, chemicals, glass, from 60% to 90%. UK industrial power prices are among the highest in the OECD and the main reason heavy manufacturing leaves. Two days later Brussels launched a €1.5bn Battery Booster, interest-free loans from carbon-market revenue, capped at €500m a project, aimed at the scale-up phase where European battery plants keep dying. It backs a real capacity wave: Tesla has doubled Giga Berlin's cell target to 18 GWh and opened the line to outside engineers, and Eni broke ground on an LFP gigafactory in Brindisi.

Both attack the right layer: the cost of energy and capital, where reindustrialisation actually dies.

One layer stays out of reach. Roughly 94% of finished rare-earth magnets are made in China, and last week a German startup, alqem, raised just €8m to design magnets without rare earths at all. Set that against Beijing's magnet export controls, suspended only to November 2026, and the €8m reads as rational insurance.

Europe can subsidise cells and power bills; the magnet in every motor, robot and turbine is still China's to switch on and off.

GOV.UK: The UK's Modern Industrial Strategy: year one

EARNINGS TO WATCH

• Kongsberg Gruppen (KOG), Monday 13 July. Q2 from a European defence prime. Watch missile and maritime-autonomy order intake against the undersea build-out above.

• AAR (AIR), Wednesday 15 July. Aftermarket parts and services demand as a read on the commercial-aviation maintenance cycle.

• TSMC (TSM), Thursday 16 July. The quarter the whole compute chain reads. Capex guidance and advanced-node and packaging utilisation set the tone for AI supply.

• ABB (ABBN), Thursday 16 July. Electrification and robotics orders. The grid-equipment backlog that gates data-centre build shows up here first.

• Intuitive Surgical (ISRG), Thursday 16 July. Procedure volumes and system placements, the cleanest read on surgical-robotics demand.

• GE Aerospace (GE), Thursday 16 July. Engine deliveries and spares margin, and shop-visit cadence on the narrowbody fleet.

• SAAB (SAAB-B), Friday 17 July. Nordic defence order book, another gauge of the European rearmament run-rate.

• Volvo Car (VOLCAR-B) and Autoliv (ALV), Friday 17 July. European auto demand and safety-content pricing into a tariff-shadowed second half.

MACRO

• China Q2 GDP and June activity data, week of 14 July. The industrial-production and fixed-asset-investment prints show whether the excavator-export strength above rests on real demand or export front-running.

• US June CPI, week of 14 July. Feeds the rate path that sets the hurdle rate on every industrial capex plan in this edition.

EVENTS

• The 16 July semis-and-robotics tri (TSMC, Intuitive Surgical, ABB) opens the industrial earnings season in earnest.

• Tesla's Giga Berlin Cell Giga Challenge closes to applicants on 24 July, the first read on whether outside engineers can move the 4680 yield curve.

ON THE RADAR

• FCC satellite-licensing vote, expected late July, now shadowed by an Earthjustice petition for a full environmental review of orbital data-centre constellations after filings for nearly 140,000 satellites. The first regulatory brake on space-based compute.

• China's rare-earth magnet export-control suspension runs to 10 November 2026. Any re-imposition signal is the switch behind the alqem bet above.

• Proxima Fusion's €411m round, Europe's largest fusion cheque, put Google and RWE on the cap table. Watch whether other hyperscalers follow European firm-power bets.

• NRC NEPA comment period runs to 21 August. The litigation posture that forms around it will set US reactor timelines for years.

I built The Brief to be the most useful 6 minutes in your week. If it is, please forward it to a friend. If it is not, send me a message on dan@danmottram.com and tell me why. You can also contact there for podcast bookings, capital raising support, and strategic advisory work.

ISI. Signal over noise.

The market spent a decade calling Planet Labs a bad business with a beautiful mission. It photographs the entire land surface of the Earth every day, whether or not anyone asks, and it lost money doing that from 2010 to 2025. The imagery looked like a commodity. The losses were real.

Then two things happened in one year. Planet turned its first operating profit, and its founder stopped calling the fifteen-year pile of images an archive and started calling it a model.

The satellites did not change. What the pictures are for did.

Key insights

• The category was wrong: Planet was priced as a satellite company, and the asset is a fifteen-year daily index of the Earth that no rival can rebuild.

• AI reprices the archive: the same record that made Planet look uneconomic is the training substrate for models that answer physical-world questions.

• The tipping point is a number: orbital compute flips when launch cost enters a $200-500/kg band, and the space-AI bet rides on when that arrives.

Planet crossed into adjusted-EBITDA profit in FY2026, the first time in its history, as revenue growth reaccelerated to 26%. The FY27 column is the guidance midpoint.

Subscribe now

A satellite company, supposedly

Planet was built to be underestimated.

Three NASA Ames scientists incorporated it in 2010 on one idea: fly cheap smartphone-grade electronics in place of exquisite satellites, and iterate the hardware like software. They called it agile aerospace. The satellites are shoebox-sized Doves. Roughly 462 have launched, about 200 fly now, and since November 2017 they have imaged the planet's whole landmass every day. It listed through a SPAC in 2021 at a $2.8bn valuation.

Then it spent two years being told the habit of photographing everything was a cost with no market under it. For years the market saw a commodity camera in orbit and priced it as one.

What "data" actually means here

Call it "data" and you learn nothing.

Data is not an asset on its own. A container of someone else's telemetry is worth nothing to anyone who cannot use it, and most data ages into noise. The words "satellite company" are just as empty. They group Planet with operators who own sharper cameras and none of what counts.

What counts is narrower. Planet owns a single, consistent, daily record of the same planet, taken the same way, for fifteen years. The value is not in any image. It is in the run: how a field, a port or a front line changed yesterday against every comparable day before.

That record cannot be bought late.

A rival with more money can out-resolve Planet tomorrow. It cannot photograph 2017. The moat is made of time, and time does not respond to capital.

Planet's own line is the sharpest version of it. "Google indexed the internet to make it searchable. Planet's indexing the Earth to make it searchable."

From pixels to answers

The index needed a reader.

Selling raw imagery meant selling to the few customers who paid analysts to stare at it. A large model reads the whole archive at machine speed. Ask what a specific flood looks like today, set against every flood the system has seen, and get an answer and a report with no analyst in the loop.

That is a different business. Selling answers is higher-margin and stickier than selling images by the square kilometre. Planet spent a decade buying the analytics to sit on top, from RapidEye's archive to Sinergise's platform. It added sensing the Doves cannot: high-resolution Pelican tasking with onboard compute, and Tanager's hyperspectral eyes that read methane plumes facility by facility. The model layer is a partnership: Anthropic's Claude signed in 2025, an Nvidia engine in 2026 that cuts raw imagery to analysis in seconds.

The pixels were never the product. The index is.

Three sensing layers feed one queryable index. The Large Earth Model turns the archive into answers a user can query in plain language.

Why compute moves to orbit

The next step follows from a bottleneck.

A constellation that photographs everything produces far more than it can send down. So Planet is starting to compute in orbit, next to the sensor, and transmit answers where it used to transmit pixels. It already runs Nvidia chips in space, and it is Google's satellite partner on Project Suncatcher, which flies TPUs in orbit. The first two-satellite demo is due in early 2027.

The case rests on one number. Compute is an energy problem, and orbit is a good place to solve it: roughly five times the solar power, passive cooling, no competition for land or grid. Set against that is the cost of getting mass to orbit.

Today that cost is around $1,500/kg to low orbit. Orbital compute flips when it reaches $200-500/kg.

Launch cost must fall into the $200-500/kg band, about 3-7.5x below today, before orbital compute beats the ground. Planet dates the crossing near 2029, Google nearer 2035.

The claimants disagree on the date more than the number. Planet puts the crossing near 2029, Google's own Suncatcher paper nearer 2035. Six years is a long time to leave in a business case.

The one open question

Strip away the noise and one thing is genuinely unknown.

The archive is real, the profit is real, the partnerships are signed. What no one knows yet is who captures the value when data meets models: the company that owns the fifteen-year record, or the companies that own the models reading it. Planet is betting on the record, because models are being commoditised and the archive cannot be rebuilt.

The near-term business does not depend on the answer.

Government pays for it now. Planet holds imagery work with the US National Geospatial-Intelligence Agency and a seat on the National Reconnaissance Office's commercial layer. In July 2025 it signed a €240m deal with Germany for dedicated access to its high-resolution network. The backlog reached $900m, up 79%, and roughly 97% of revenue recurs. The top line grew 26% last year to $308m, and 42% last quarter.

One caveat sits under the profit. Planet's reported FY2026 result is still a large net loss on paper, most of it a non-cash charge that grows as the share price rises. The operating signals point the other way: positive adjusted earnings, positive free cash flow, an accelerating top line.

If Suncatcher slips, Planet keeps a profitable data business and loses an option.

My read: Planet spent fifteen years being judged as a satellite company and priced as a bad one. The satellites were always the cheap part. It was quietly building the one dataset the AI era cannot manufacture on demand, and it reached profit at the moment that record turned from sunk cost into scarce input.

The question is not whether Planet takes better pictures. It is whether owning the memory of the Earth beats owning the models that read it.

Subscribe now

On 30 June the government published the Defence Investment Plan: £298bn over four years, with £15bn of genuinely new money layered on top of the existing baseline.

£5bn goes to drones and autonomous systems, the largest drone commitment the country has made.

£11bn goes to munitions and stockpiles, with a commitment to at least six new energetics factories, the plants that make explosives and propellant.

£3.2bn goes to space. The Royal Navy takes hybrid drone-carrier command ships in place of part of the destroyer line it had been promised.

Is £298bn a lot? Mostly it is the old trajectory rebadged. At around £75bn a year the UK sits near 2.6% of GDP against a stated path to 3.5% by 2035, and the plan does not close that gap.

The honest new steer is the £5bn, the £11bn and the six plants. So the better question is what that composition buys. Defence money buys three different things: platforms that fight, consumables they fire, and the capacity to make more of both.

A destroyer is the first. A shell is the second. An energetics factory is the third, and it is the one Whitehall has decided it cannot do without.

Nobody sinks capital into precursor-chemical capacity unless they expect to consume at a rate that empties the shelves. Starmer cited Ukraine burning through roughly 200,000 drones a month. The factories are the same sentence written in concrete.

There is a second decision hiding under the first. A destroyer programme and a drone-and-energetics programme are built by different industrial bases. Warship money flows to the same prime yards it has flowed to for a century. Drone and munitions money flows to chemical plants, machine shops and firms that did not exist five years ago. Shifting the mix shifts the supplier base, and the incumbents know it.

My read: the factories are the condition of entry, and a condition of entry is what this plan buys. Standing up munitions capacity takes three to five years and runs on scarce process-chemistry labour, so the constraint is people and permits before it is money. Britain has diagnosed the next fight as consumption-limited, which is right.

Whether the change of Prime Minister inside the plan's own four-year window keeps the drone-over-destroyer bet is the part that time will tell. The number that decides whether this plan mattered is the “spade in the ground” date of the first energetics factory, and it is not in the document.

GOV.UK: The Defence Investment Plan

On 1 July HII's Ingalls yard began fabrication of the future USS John F. Lehman, a Flight III destroyer. The ship is routine.

HII is fabricating the major structural units across six partner yards in Texas, Louisiana, Mississippi and Florida, then shipping the steel to Pascagoula for final assembly and integration. It expects to outsource more than 2.5m shipbuilding hours in 2026.

A warship build has always been one integrated job at one site. HII has split it into portable work and prime-yard work, because only systems integration and sea trials have to happen at Pascagoula.

The reason is welders.

The Gulf Coast catchment cannot hire them fast enough. The dollars for the backlog of five destroyers under construction, with seven more behind, are already appropriated; the labour hours to spend them are the gate. Distributed fabrication converts a fixed-site shortage into a national one, and pulls second-tier yards into the naval base as it goes.

Two things follow. Structural steel commoditises out to whoever has spare bays and a workforce, while the combat system and the trials, where the value sits, stay with the integrator. And units built at yards new to naval work must hit naval standard, so the seams are where this either proves itself or fails. Whether this is a temporary workaround for a labour market or the permanent shape of American shipbuilding depends on the rework rate, and nobody knows it yet.

Naval News: HII Begins Fabrication of Flight III Destroyer John F. Lehman

Hold the two shipyards side by side.

HII is re-plumbing one destroyer across six yards to get around a welder shortage. In the same quarter, official Chinese data put China at 84.9% of global new ship orders by deadweight tonnage, 59.53m DWT.

South Korea took 12.8% and Japan 1.4%. China held 57% of completions, led new orders in fifteen of the eighteen major ship types, and holds about 70% of the global order backlog. A book at 70% means Chinese yards are sold out for years, and a shipowner who wants tonnage soon has nowhere else to go. That is a moat built from berths and welders, the same two inputs America is scrambling to redistribute.

The only Western brake is a policy switch currently set to off. The US finalised Section 301 port fees on Chinese-built and operated ships in 2025, then suspended them to November 2026 under the trade truce. The fees say Chinese shipbuilding dominance is a strategic exposure. The suspension says American trade cannot yet afford to tax the ships it runs on. Both are true, and the second is winning.

Forward Prediction: the Section 301 port-fee regime lapses or is bargained down again at the November 2026 expiry, and does not resume at full rate.

SCMP: China tightens its grip on global shipbuilding, grabbing 85% of new orders

On 30 June BMW completed a $1.7bn investment in South Carolina, the Spartanburg expansion plus a battery-assembly plant at Woodruff, and premiered the new X5.

Spartanburg can now build one model across five drivetrains on a single line: petrol, diesel, plug-in hybrid, battery-electric and hydrogen. The iX5 becomes the first fully electric BMW assembled in the US from late 2026.

A single-powertrain electric plant is the optimised bet: cheaper per unit, stranded the moment demand softens, the fate that already shelved two of Europe's ACC gigafactories.

A five-powertrain line is the flexible bet (optionality): it gives up some line efficiency to keep the option of shifting the mix year by year without re-tooling. Car makers have argued about optimised versus flexible since the first dedicated EV platforms, and BMW has just paid $1.7bn to stay on the flexible side of it. Against a plant already turning out more than 400,000 units a year, that is the price of flexing an existing asset. Keeping hydrogen and diesel on the same line as battery says BMW expects no clean, dated cutover, anywhere, for years.

Around half of Spartanburg's output ships abroad, so the tariff regime and the demand mix jointly set the payoff, and localising pack assembly at Woodruff answers the content rules directly.

BMW Group: BMW completes $1.7 billion U.S. investment and premieres the new X5

Shenzhen's X Square Robot said this week that four consecutive rounds, ending in a Series C, carried its valuation past $2.8bn, with IDG, HongShan and Xiaomi among the backers.

The money funds an architectural dissent. Most embodied-AI firms bolt together separate modules for vision, language and action. X Square's WALL-B trains perception, language, action and physical prediction inside one network, a single model of how the world behaves.

Which architecture wins decides where the value in a robot ends up.

If the modular approach wins, the model layer breaks into swappable blocks and commoditises, and the chassis makers keep the margin. If the unified world model wins, the trained network is the moat and everything below it is iron.

X Square holds the second position and is open-sourcing part of the stack to make WALL-B the thing others build on. The open-sourcing is also an admission. A lone unified model competes with the entire modular camp at once, so X Square needs that camp building on its terms.

The claim now has to survive the floor.

X Square's open release cleared four of seventeen real-robot tasks, and lab accuracy near 95% has a habit of arriving at roughly 60% in the field. China's summer robotics conferences are where WALL-B either shows field generalisation or does not. Paid volume in the cleaning and in-home pilots is the test; demonstrations are not.

The Robot Report: X Square Robot brings valuation to $2.8B in four consecutive rounds

On 30 June Beehive Industries bought two Cincinnati-area machine shops, Able Tool and Planet Products, weeks after ordering 30 EOS metal printers, the largest such order it has disclosed. The stated target: more than 8,000 of its 3D-printed Frenzy jet engines a year, small turbojets for low-cost cruise missiles and one-way drones.

Printing the engine core is the solved half. A printed core becomes a flight-ready engine through machining, finishing and inspection, and Beehive just bought that half.

Thirty printers against an 8,000-engine target implies under one engine per printer per working day; the machine shops either keep up or the printers idle. Buy the constrained step in the same news cycle as the printers and you have said what you believe: the printing works, and the race has moved to industrialisation. What remains is beating cast incumbents on unit cost at volume, and that turns on skilled machinists. Hence "120 years of combined experience" in the press release.

Additive removes the tooling and the long lead times of cast turbine parts, and in doing so it makes a new category real: a jet engine cheap enough to expend. Britain has just committed £11bn to consumables on the same logic.

BusinessWire: Beehive Industries Acquires Two Cincinnati-Area Machine Shops

On 30 June Schneider Electric agreed to buy Cognite for $3.1bn in cash and fold it into Aveva, its industrial-software arm. Cognite turns the mess of a plant's sensor tags, historians and maintenance logs into one trusted, queryable data model. Its Atlas layer lets an AI agent take action on top. Cognite runs on about $170m of revenue. Schneider, a company built on hardware, is paying roughly eighteen times sales for it.

Why pay a software multiple when the AI models themselves are commoditising?

Because Schneider is buying the layer the models cannot supply. It already owns the controllers and drives at the bottom of the stack. What it lacks is the middle: the contextualised data that lets an agent know a tag is a reactor inlet temperature, with limits and dependencies.

Context is what keeps an agent's action from breaking something. Whoever owns that layer sits between every model vendor and every piece of plant on the floor. The cost being removed is the engineer who today reconciles operational data by hand before any decision gets made. The named loser is the systems-integrator channel that bills for exactly that reconciliation.

Siemens is racing for the same layer through its NVIDIA work, so the majors have converged on the same map: the models commoditised, the contested prize is trusted industrial data. Whether that data stays a vendor-owned moat or follows the models into commodity is the open question, and it is worth a lot more than $3.1bn.

Bloomberg: Schneider to Buy Industrial AI Firm Cognite for $3.1 Billion

On 1 July Oxmiq Labs closed a $35m round led by Fundomo and the Samsung Catalyst Fund, with MediaTek and Pegatron in.

The founder is former Intel graphics chief Raja Koduri. The product is OxCore, a licensable block fusing GPU, CPU and a tensor engine, sold with a CUDA-compatible software path. The stated aim is to be the Arm of custom AI silicon.

The bet is that accelerator design de-verticalises the way processor design did once Arm split the IP from the fab. A custom-accelerator programme costs hundreds of millions and several years before first silicon. Oxmiq sells the starting point that removes most of that bill for hyperscalers, sovereign-AI efforts and second-tier chip firms. Out-designing Nvidia is nowhere in the plan. Arm earns roughly $3.2bn a year on IP it never manufactures, a slice of the whole industry's volume off a single core, and that is the shape being copied.

Walk the chain and the chokepoints are visible.

TSMC holds the fab. Nvidia holds the software, because CUDA is where fifteen years of tooling and habit live, which is why a CUDA-compatible path is the most load-bearing item in Oxmiq's spec sheet.

The architecture layer has sat captive inside each designer, and Oxmiq is trying to prise it out and sell it sideways. Foundry and ODM money in the round says the manufacturers want a licensable core to point their capacity at. Jim Keller joining the board says the silicon crowd rates the team. Whether this becomes Arm or becomes one of the many companies that tried to become Arm is not callable at $60m raised.

EE Times: Oxmiq raises $35m for GPU IP, expands focus to data center design

National Grid Ventures agreed around 1 July to put $1.75bn into a 35% stake in the US developer Joulent, whose first project is a 2.67GW gas plant in West Texas.

The plant is built 50/50 with Chevron and feeds a Microsoft data-centre campus under a twenty-year contract, turbine to rack, off the wider network.

A British transmission utility has decided the fastest route to AI-power revenue runs outside the regulated grid it owns. The scarce asset is a timely, firm connection; this deal is a product for skipping the queue, sold by a company that owns the queue. I covered where the data-centre bottleneck went in the Helix article. This is the same constraint attacked from the supply side.

A Duke Energy subsidiary agreed around 29 June to sell its last North Carolina offshore-wind lease back to Interior for about $129m, against the $155m it paid in 2022.

Duke says the money refocuses on firm generation, possibly nuclear and gas. It is the ninth offshore-wind project unwound with the federal government since the administration took office, and on Duke's balance sheet it is a rounding error.

The denominator behind it is not.

Wood Mackenzie counts roughly $121bn of early-stage projects, about 92GW of solar, wind and storage. All of it is now routed through Interior review at the Secretary's own office, above the career staff who used to sign it off. A further 29GW of onshore wind sits held by the Pentagon. A federal judge issued an injunction in April; Interior appealed on 17 June. An administration that spends legal capital to keep a permitting queue closed is telling you the queue is the policy.

Name the mechanism and the beneficiaries fall out.

Permit delay keeps renewables out, and keeping them out keeps firm gas and nuclear in. It is an industrial subsidy to dispatchable power, paid through the permitting queue, and whoever already holds permitted firm capacity is repriced upward every time the pipeline slips. Duke selling a wind lease to fund gas and nuclear is a utility reading that policy correctly.

Utility Dive: Duke gives up offshore wind lease for 'partial' reimbursement under Interior deal

On 25 June China's NDRC and NEA released the 15th Five-Year Plan for the energy system, running to 2030. China had already installed about 1,234GW of solar and 650GW of wind by February. That is more solar alone than the entire US generation base across all sources. It added roughly 264GW in the first half of last year. The nuclear track targets about 110GWe by 2030, close to double the current US operating fleet, built on a standardised Gen-III reactor while the US runs a menu of pilots.

Read the plan for its assumptions: the binding input is manufacturing throughput, and China acts as if it has already won that contest. The volume learning-curve accrues to the Chinese EPC, reactor and grid-equipment chains, which threatens Western equipment makers even where tariffs wall off their home markets. The permitting fight in the item above and this plan answer the same question: what decides what gets built. The two systems gave opposite answers. One gates on approval, the other on absorption.

Absorption is now the only constraint left. Whether transmission and curtailment can keep pace with more than 260GW of renewable additions a year is the open question. China's first-half 2026 install data, due late July, will show whether the pace held.

CSIS: China's Nuclear Energy Priorities Under Its 15th Five-Year Plan

Helion, the fusion company that reached a $15.5bn valuation in a June round, cleared a gate this month that has nothing to do with plasma. Washington State issued its Orion plant a radioactive materials licence and an air emissions licence, the first fusion operating licences granted anywhere. Helion has not produced a watt of fusion electricity. Nobody has.

What it now holds is the permitting tranche of its 2028 Microsoft delivery deadline, signed years ahead of the physics, and that order of operations is rational.

The binding constraint on a fusion contract was never only the reactor. It was the stack of non-physics gates around it, permits, interconnection, safety approval, each with its own multi-year queue.

Helion has cleared the regulatory one first. Meanwhile the demand side widens.

OpenAI is reported to be discussing Helion supply scaling to 5GW by 2030 and 50GW by 2035, a hundred times the 50MW Microsoft anchor. That reported appetite is the reason Sam Altman gave up the chair. Two hyperscalers are now positioning for electrons a fusion plant has yet to make. What is being priced is firm future power. The reactor is the option attached to it.

World Nuclear News: Fusion plant clears regulatory milestone

EARNINGS TO WATCH

• Yaskawa Electric (6506.T), Friday 10 July, Q1 FY2026. The core automation bellwether and the first clean read on the robotics order tape this quarter. Watch order intake and the FY2026 operating-margin guide after a soft prior year.

• Enerpac Tool Group (EPAC), release Tuesday 7 July, call Wednesday 8 July, Q3 FY2026. Short-cycle industrial tools; a read on maintenance and MRO capex in the US and Europe.

• AZZ (AZZ), release Wednesday 8 July, call Thursday 9 July, Q1 FY2027. Hot-dip galvanising, the rust-proofing step for structural steel, plus electrical products; a proxy for US steel, infrastructure and reshoring construction demand.

MACRO

• FOMC (Federal Open Market Committee) June minutes, Wednesday 8 July. The rate path sets the discount rate on the entire capex cycle: grid, fabs, defence and data centres are all long-duration bets sensitive to where the Fed signals it lands.

• China CPI and PPI, around Wednesday 9 July. PPI is the cleanest read on whether the China price war in autos, solar and robotics is still exporting disinflation into Western supply chains.

• UK monthly GDP and industrial production, around Friday 11 July. The manufacturing base the new defence-industrial spend has to lift from.

EVENTS

• US Independence Day spillover, week open. Thin US flow into Monday 7 July after the 4 July holiday; expect a quiet start.

• Korea Gwangju fab site decision, reported as near. The trigger event behind Samsung's public call for new nuclear to feed a 6.3GW chip cluster.

• China 15th Five-Year Plan follow-through. Watch NDRC and NEA for nuclear-newbuild and grid implementation notices landing in the following weeks.

• The 24 July tariff expiry. The Section 122 flat 10% tariff lapses automatically on 24 July, and USTR is expected to roll targeted countries into Section 301 tariffs, 10 to 12.5% with no statutory cap, with investigations concluding around 20 July. A first-order catalyst for autos, machinery and metals input costs, two weeks out.

• A 250-ship Senate bill. Legislation introduced 3 July targets 250 new vessels over ten years. If it moves, it is demand-side fuel for exactly the labour-gated capacity problem in this week's HII item.

• Oklo Groves first criticality, targeted for July. The DOE cleared the final safety analysis around 1 July; criticality in-window would be the first commercial-fuel advanced reactor to reach it on private land.

• Interior permitting appeal. Any appellate movement on the clean-energy injunction reprices the 92GW pipeline and every firm-power name against it.

• Grid-equipment order flow. After Hitachi Energy's Virginia transformer factory and National Grid–Joulent, further transformer, turbine or off-grid data-centre deals keep the supply-side constraint moving.

  Subscribe now

I built The Brief to be the highest-return minutes of your week.

If it is, please forward it to a friend or peer.

If it is not, please tell me why via message here, or at Dan@DanMottram.com.

Same address for podcast bookings, collaborations, and strategic advisory work.

ISI. Signal over noise.

Subscribe now

The figure is not a fresh commitment. It completes a plan announced in 2022, covering the Spartanburg expansion and the new battery plant at nearby Woodruff. Read against the roughly US$16bn BMW has put into the site since 1992, the headline number is a final instalment. It closes an old plan.

The industrial signal sits in what the money buys.

Spartanburg will build the new X5 across five drivetrains on a single line: petrol and diesel mild hybrids, a plug-in hybrid, the fully electric iX5, and later a hydrogen fuel-cell variant.

The iX5 becomes the first fully electric BMW assembled on US soil, with output starting before the end of 2026. High-voltage batteries come from Woodruff using cell-to-pack assembly, which drops the module stage and shortens the pack line.

The reason this matters is Spartanburg's role as an export machine.

The plant supplies global demand well beyond the domestic market. It is BMW's largest anywhere by volume, producing 412,799 units in 2025 and shipping close to half of them abroad. It was the largest US automotive exporter by value last year at $9bn. Adding EV and battery capacity to a plant of that export weight converts a combustion-era export hub into an electric one, with no greenfield site and no stranded tooling.

That is the localisation case in its cleanest form. BMW is deepening an already-localised US base at the one plant where domestic build already feeds global demand. A single flexible line hedges the drivetrain-mix uncertainty that has stalled dedicated EV plants elsewhere.

Of course, we then need to ask the “so what” question.

The competitive edge in US EV manufacturing looks to be shifting from who builds the biggest dedicated EV plant to who can convert high-volume export capacity without losing it.

Spartanburg is the experiment. Watch iX5 export share once the line ramps: that number will show whether US-built electric BMWs travel as well as the combustion X models did.

Subscribe now

Sources: BMW Group press releases (Plant Spartanburg "Home of X", 30 June 2026; largest US automotive exporter by value, 2025); Fox Business; Automotive Manufacturing Solutions.

It surfaced when KR Sridhar, founder of Bloom Energy, sat down with on the podcast, and the lesson has almost nothing to do with energy.

Subscribe now

Sridhar built Bloom on solid-oxide fuel cells in the early 2000s, on technology that traced back to generating oxygen for a Mars mission.

For most of two decades the company was slow, expensive and much-doubted, selling on-site power to customers who mostly did not need it yet.

Then the AI build-out arrived, the grid ran out of headroom, and the thing he had been building became the thing everyone suddenly wanted. Overnight, by the standards of a twenty-year wait.

It is tempting to file this under prediction. He saw AI coming. He did not.

Nobody in 2001 modelled hyperscale data centres pulling gigawatts through a grid that could not keep up. What he did was different, and more worth copying.

He bet on a vector, not an event.

The vector was a direction of travel. Power would become more distributed, more local, and valued more for reliability and speed than for headline cost. That direction held across a wide range of futures.

It did not depend on AI. It would have paid off on a longer or shorter fuse under electric vehicles, under reshored manufacturing, under climate volatility, under any world that strained a grid built for a slower age. He never had to name the trigger. He needed the direction to be right.

This is the first move. Prediction picks an event and a date. Positioning picks a direction and waits. The forecaster who says “AI will need this much power by 2027” is playing a game they will probably lose on the specifics. The operator makes a quieter claim. Power will get scarcer, more local, more time-sensitive, and I want to be standing there when it does. The exact path matters far less to that bet.

When Stebbings pressed him on the bubble question, Sridhar called the demand a hockey stick on a hockey stick. He was describing a direction that had finally turned vertical, not a number he had forecast in advance.

The second move is harder, and it is where most foresight dies. Being early is indistinguishable from being wrong, right up until the moment it is not.

For twenty years the company looked like a bet that had not come in. The market’s verdict on an early position and a mistaken one reads the same: no demand, burning cash, a story that has not landed. The difference only shows up later, and only if you are still there to see it. Most visionary failures were correct about the direction. They were under-capitalised for the wait. Surviving the gap between the bet and the payoff is the actual skill. The idea is the easy part.

The third move is knowing where to stand. In any system, when you relieve the bottleneck everyone is watching, the binding constraint moves. It rarely announces itself. It slides to the least glamorous layer, the one no one is optimising. In the AI story the industry poured its attention into chips and models. The constraint slid underneath, to power, to permits, to transformers with four-year lead times. Sridhar’s advantage was not calling the winner at the top of the stack. He had parked, years earlier, on the layer the constraint would eventually reach.

You can run this as a question.

Where is the bottleneck now, what happens when it clears, and what becomes the next limit once it does. The answer points at where to position. It is usually somewhere boring, slow to build, and easy to ignore while the exciting layer soaks up the capital and the headlines.

Put the three together and foresight stops being mystical. Bet on a direction that holds across many futures. Park on the constraint that the direction will eventually expose. Build something that can survive being early, because early is where you will live for a long time.

Most of us cannot wait twenty years, and most bets should not need to.

The fuse can be shorter. The structure holds at any length. A career, a product line, a company can each be positioned this way. Choose the trend you would still back if the specific catalyst never showed up. Find the input that everyone treats as free and infinite today, and ask what happens when it is not. Then make sure you can afford to be right slowly.

The trap is the opposite of what people assume. The danger is rarely that the direction is wrong. It is that the payoff arrives after you have run out of road. Foresight without endurance is an expensive way to be right for someone else’s benefit.

My takeaway: the operators who compound take positions that do not need luck on timing. They build enough slack to still be holding when the world finally agrees.

Subscribe now

Link to the show below, it’s worth a watch, let me know what you think:

It builds rockets, it builds satellites, and it flies both for the companies that own the businesses sitting on top. On 29 June it agreed to change that.

Rocket Lab will buy Iridium Communications for 54 dollars a share, an enterprise value near $8b. In the process it stops being a supplier to the space economy and becomes an operator inside it.

Subscribe now

The price tells you what kind of deal this is.

Iridium traded at about 53 dollars before the announcement, close to its 52-week high. Rocket Lab is paying up, half in cash at 27 dollars a share and half in stock. A 3.6 billion dollar bridge facility from Deutsche Bank and Wells Fargo funds it. Rocket Lab is paying a full price because it wants what Iridium owns.

What Iridium owns is the interesting part.

Sixty-six satellites in low Earth orbit, cross-linked so they route traffic between themselves, which gives genuine pole-to-pole coverage that rivals cannot assemble quickly. It is the only mobile satellite network with truly global reach.

Underneath the satellites sits a scarcer asset again, globally coordinated L-band spectrum, licensed and hard to challenge. On top sit 2.5 million paying subscribers.

There is a long-running US Department of Defense contract for sovereign mobile communications, and safety-of-life services that aircraft and ships depend on.

The numbers describe a utility. In 2025 Iridium turned over $872m, produced $495m of operational EBITDA at a 57% margin, and converted roughly $296m into free cash flow. Recurring, high-margin and cash-generative. A launch and manufacturing business earns the opposite way, in lumpy projects tied to individual contracts.

That is the logic of the deal.

Rocket Lab is buying its way from the bottom of the space stack to the top.

It designs, it builds, it launches, and now it will own and operate the network those capabilities serve. It gains a captive customer for its own rockets and satellites. It gains a spectrum position that takes decades and a regulator's blessing to build. And it gains defence relationships that are close to impossible to acquire any other way.

The market read the deal as a shot at Starlink. That framing is worth resisting.

Iridium runs narrowband L-band services, voice, tracking, machine-to-machine data, direct-to-device messaging and safety communications. It does not sell broadband internet, and it will not be racing SpaceX for home connections. Craig-Hallum made the point plainly, and William Blair moved Iridium to a neutral view once the agreed price capped the shares.

My read: the value here has little to do with beating Starlink at anything. Rocket Lab has turned itself from a business that sells shovels into a business that owns a mine. The recurring cash from the mine keeps the shovel factory running. That is a better position than being the best supplier in a gold rush.

The pattern travels beyond space. When your product is a capability that other people rent, the durable margin sits one layer up, in the service that capability makes possible. Owning the operating layer, and the spectrum behind it, is where a collection of separate space businesses turns into a single stack. Rocket Lab has just built the latest example of that move.

Two things still have to hold.

The debt is real, a 3.6 billion dollar bridge on top of the stock consideration, and it will need refinancing on decent terms. The deal does not close until the middle of 2027. It must clear an Iridium shareholder vote and a wall of regulators. They will study the spectrum transfer, the defence contract and a foreign buyer taking over sovereign communications. Neither risk is trivial. Both are the price of buying a decade of build in a single step.

If it holds, Rocket Lab stops being the most interesting supplier in space and becomes one of its few complete companies.

Key insights

• Castelion is in talks to raise at about $12bn, roughly four times its $2.8bn December mark. The re-rate tracks current investor appetite for defence manufacturing.

• Its Blackbeard missile targets a unit cost of about $300,000 to $440,000. The Army's Dark Eagle hypersonic runs near $37m a round. At $37m a hypersonic is held in reserve; at $370,000 it is something you can spend freely. At one-hundredth the cost, the doctrine you can afford changes.

• The method is SpaceX's: vertical integration, build-and-fly testing, design-to-cost. Castelion reports more than 20 flight tests in 2025 and a first flight about 18 months from a clean sheet.

• The US Navy placed a production order in June 2026, $23.4m for 50 prototype rounds, after roughly $155m of development awards earlier in the year. A Pentagon framework agreement targets a floor of 500 missiles a year once the design is validated.

• The bet rests on production rate. A solid rocket motor line at thousands per year is the constraint, and the ramp figures behind the valuation are budget projections, not signed orders.

  Subscribe now

100x Cost Reduction

Castelion is in talks to raise at about $12bn, 4x the $2.8bn mark it set in December.

Read as a financing story, that is investor appetite for defence. The more useful reading runs the other way. The valuation is roughly what you get when you take the cost of a hypersonic missile and divide it by a hundred.

Castelion builds Blackbeard, a strike weapon it says it can make for a few hundred thousand dollars a round. The Army's Dark Eagle costs near $37m.

Price decides what a missile is for.

At $37m it is a strategic asset, counted in dozens, signed out by senior command, fired only when the target is worth more than the round. At $370,000 it is ammunition, fielded in thousands and spent on targets that were never worth a $37m shot. Same speed, same range, the same physics. What moves is the doctrine you can afford to run. That hundredfold gap is the product Castelion is selling.

The doctrine behind the money is magazine depth. A peer fight in the Pacific is partly an arithmetic problem: how many shots each side can take before the shelves empty. Hundreds of exquisite missiles can be absorbed. Thousands of cheap ones change the sum.

Castelion's framework agreement with the Pentagon targets a floor of 500 Blackbeards a year once the design is validated, with a stated path to thousands. Project Ranger, a planned 1,000-acre plant in New Mexico, is sized for that rate.

The cost claim feels credible because of how the company builds.

The three founders ran sales, operations and finance at SpaceX, and they brought the method with them. Castelion makes its own solid rocket motors, seekers, flight computers and thermal protection in house, and it tests by flying. The company reports more than 20 flight tests in 2025, and a first flight about 18 months from a blank sheet. Build, fly, break, iterate, on a cycle measured in weeks.

The low price is designed in from the first drawing, not found later.

The design is the easy half. The line is the constraint.

A flight test proves a missile works. It does not prove you can build it at thousands a year. Solid rocket motors are the hard part, and the US motor supply base is concentrated and already short of capacity. Whoever controls that bottleneck controls the ramp.

The procurement signal is already moving.

The Navy awarded Castelion about $50m in February and $105m in April, then placed a $23.4m production order in June for 50 prototype rounds. Across the same budget, the total US request for hypersonics research fell from $6.9bn in FY2025 to $3.9bn in FY2026. Money is leaving exquisite development and moving towards systems that can be built at rate.

Watch the capital, not the press release: that is the real thesis.

Run the saving backwards and the threat appears. The $37m was a cost, and it was also a wall. Exquisite, low-volume, cost-plus production is what kept the missile business to a handful of primes. Take the unit cost down by two orders of magnitude and the wall that protected them comes down with it. Cheaper missiles are good news for the buyer and a direct hit on the incumbent whose advantage was the price of getting in.

One variable governs the rest, and it is production rate. Everything else is settled enough to model. The ramp figures behind a $12bn valuation are budget-document projections. A budget line is a small knowable number standing in for a large unknown: the factory that does not yet run at rate. An appropriation can be cut. A motor line can stall. The valuation prices the plant as though it already runs.

The same move keeps working, and only the target changes.

Take a product that is exquisite, built in small numbers and priced cost-plus, and beat it on the factory floor instead of the spec sheet. Launch went that way under SpaceX. Defence Autonomy is going that way now under Anduril. Munitions are next, and Castelion is the firm holding the drawings.

My read: the hard technology risk is largely behind them, and the open question is purely industrial scale. Can they pour concrete and motors fast enough to fill a number the market has already paid for. That is a narrower and more familiar risk than it looks, and it is the same risk SpaceX cleared once.

For any operator whose moat is a high-cost, low-volume product, the question is the same. What does the version at one-hundredth the cost look like, and who is already building it?

Micron's record quarter is a claim, and the claim has failed twice before

Micron just had the best quarter in the history of the memory business. Revenue of $41.5bn, up 346% in a year. A gross margin of 84.6%, which no memory maker has ever held through a normal cycle. The guide is $50bn next quarter.

Here is the thing about records in memory. They are usually the sound a cycle makes at the top.

Subscribe now

So treat the print as what it actually is, a claim. The claim is that the cycle will not break this time. Memory has believed that twice in the past decade. In 2019, after the cloud build-out that was meant to end boom-and-bust, DRAM contract prices halved inside a year. In 2023, after data centre demand was supposed to have set a floor, Micron's gross margin fell below minus 30% and it wrote off $1.8bn of stock. Every glut in this industry began at exactly this point, the moment of maximum confidence.

So the only question worth asking about 84.6% is what is genuinely different now. Three things are.

One, the contract. Micron signed sixteen long-term customer agreements this quarter, fourteen of them locking in close to $100bn of minimum-price revenue, with floors and deposits attached. Spot pricing set the last two cycles. Take-or-pay is setting this one.

Two, the physics. A gigabyte of high-bandwidth memory eats roughly three times the wafer area of an ordinary DDR5 chip. So when the three firms that make 90% of the world's memory pour capacity into HBM, the commodity supply does not sit beside it. It shrinks. AI demand and commodity scarcity have become the same event.

Three, who is asking. In the last two cycles the sellers chased the buyers. This quarter the buyers came early and asked to be locked in. A hyperscaler that pre-commits two years of memory has stopped treating it as a purchase. It is treating it as a supply it cannot be caught without.

None of that repeals the capital race. Micron will spend about $27bn this year, and Samsung and SK Hynix are spending to keep pace. The floor is real for the bits under contract. It is soft for everything else.

That is the part the headline buries.

My read: This is not the top of a cycle or the bottom of one. It is the moment memory stopped being one market and quietly became two. The contracted tier, sold to the clouds, holds through 2026.

The spot tier turns first. It drags the carmaker, the industrial OEM and the PC builder with it. All of them are already paying 60% to 95% more in a quarter for chips they used to order on demand.

The AI memory boom lifts Micron and the hyperscalers. It exports a shortage to everyone else. Build anything with memory in it and stay off a strategic agreement, and your cost base is now set by people who are on one. They are not optimising for you.

Watch the spread between the two tiers. It will tell you more than the margin ever does.

Micron: FQ3 FY2026 results

ACROSS THE SECTORS

(ISI 04) ROBITIS & EMBODIED AI

The humanoid demo is over, and the invoice has arrived

For three years humanoid robots have been a YouTube genre.

A machine folds a shirt, a room applauds, a valuation goes up.

Agility Robotics just did the unglamorous thing. It filed to go public.

Agility is merging with Churchill Capital Corp XI, raising $620m at a $2.5bn pre-money valuation, with Foxconn leading a $200m PIPE.

The raise is the headline. The order book is the story.

Agility has booked more than $300m of multi-year orders for its Digit robot, from Amazon, GXO, Schaeffler and Toyota, with thirty-plus more customers in the pipeline. Digit has already logged over 65,000 operating hours and moved more than 100,000 totes at GXO during 2025.

This is what the demos never showed. The hard problem in humanoids stopped being the walking some time ago. It is building the machine at volume, at a cost that clears, against orders that someone has actually signed.

Agility's RoboFab plant is built for 10,000 units a year. The float pays for the factory, and the public market will now price Digit on what comes off that line.

Demos sell the dream. Invoices sell the company.

Forward Prediction: The next humanoid maker to reach public markets does so on the strength of a signed order book, within twelve months. A unit-cost claim will not be what earns the listing.

(ISI 05) ADVANCED MANUFACTURING

The AI build-out has started eating its own supply chain

That is what changed hands in industrial-manufacturing M&A over the past year, a record, up 28%, on PwC's count. A number that size usually means the tourists have arrived. This one means the opposite.

Look at what got bought.

Power equipment. Thermal management. Automation. Advanced components. Strip the labels and it is one thing, the supply base for AI data centres, the grid and defence.

The same capacity everyone else is racing to contract, being bought outright.

The concentration confirms it. Deals above $5bn made up 56% of the value, and US-targeted deals nearly doubled to $72bn. PwC counted 155 convergence deals, the ones where a buyer swallows a firm that feeds its own supply chain.

You do not buy your supplier to flip it. You buy it because you have decided the shortage is structural and you want to own the chokepoint before a competitor does. This is what conviction looks like on a balance sheet. It also quietly removes a supplier from everyone downstream who assumed it would stay independent.

Source: Industrial manufacturing M&A hit record $173B over past year: PwC

(ISI 08) DATA CENTRES & INFRASTRUCTURE

Qualcomm books a 2029 number against a 2028 chip

Announcing a $15bn business is the easy part. Qualcomm did it this week.

It is entering AI data centre CPUs and inference against Nvidia, Intel and AMD, and it has signed two hyperscalers, one of them Meta. The target is $15bn of annual data centre revenue by fiscal 2029, with each hyperscaler deal expected to clear $1bn within a year of end-2026.

Then you read the roadmap and the dates stop cooperating.

The accelerators arrive in steps, AI200 in 2026, AI250 sampling mid-2027, AI300 sampling 2028. The Dragonfly C1000, the 250-core CPU the whole plan rests on, does not reach commercial availability until 2028. The product that has to earn the money turns up about a year before the money is due.

Forget whether the silicon is good. Qualcomm builds good silicon.

The real question is whether a mobile-chip company can take share in a market three incumbents already own. It has to do it from a standing start, on a clock where the flagship part lands in the same breath as the revenue. The $15bn and the per-deal $1bn are company targets. The next two roadmap dates decide whether 2029 is a plan or a press release. Right now it is a press release with two customers attached. That is a real start. A business is the next thing it has to become.

(09) ENERGY, NUCLEAR & GRID

Public credit is aiming straight at the nuclear bottleneck

For twenty years the story on US nuclear was that nobody wanted to build it.

That was never quite the truth. The harder fact is that nobody could forge the parts.

The US Department of Energy is putting $17.5bn of low-interest credit into the nuclear supply chain. It is routed through five Westinghouse special purpose vehicles (SPVs), each of which has to raise $1bn of private capital.

The money is for components, the heavy long-lead forgings that have been the real constraint for two decades: pressure vessels, steam generators and coolant pumps. Whole plants are somebody else's bill.

The AP1000 design has been settled for years.

The chokepoint was always the foundry, and the credit is pointed straight at it, with a target of ten large reactors under construction by 2030.

Then the part nobody joined up. Westinghouse is co-owned by Cameco, the Canadian uranium producer. The same week, Canada published its first national nuclear strategy, promising new reactors and expanded uranium output as it bids to become an energy superpower. The supply chain America is funding runs through Canada. Two governments aimed capital at the same physical bottleneck in the same week, from opposite sides of one border.

Call that what it is, industrial policy converging on a single agreed constraint.

SOURCES:

Latitude Media: DOE is dedicating $17.5 billion in loans to the nuclear supply chain

World Nuclear News: Canada unveils nuclear plans in government strategy

THE WEEK AHEAD

EARNINGS TO WATCH

• AeroVironment (AVAV), Monday 29 June, Q4 FY2026. The clearest public read on uncrewed-systems and loitering-munition demand. Watch the backlog and the book-to-bill ahead of the headline beat.

• Lotus Technology (LOT), Thursday 2 July, Q1 2026. Deliveries and cash burn are the live questions for a premium EV maker selling into a price war.

• Yaskawa (6506.T / YASKY), Friday 3 July, Q2. A pure barometer for factory automation. Servo and robot order trends, and China demand in particular, set the tone for the automation cohort.

MACRO

• US ISM Manufacturing PMI (June), Wednesday 1 July. The cleanest monthly read on US factory activity, and the first hard data point on whether the capex cycle these stories describe is showing up in output yet.

ON THE RADAR

• The SpaceX IPO. The loudest industrial story of the moment. I’m staying off the hype and watching the print and the Starlink-vs-Launch-vs-Hyperscaler revenue split. The underwriter queue is noise.

I built The Brief to be the most useful six minutes in your week.

If it is, forward it. If it is not, reply to dan@danmottram.com and tell me why.

Same address for podcasts, industrial fundraising, and advisory work.

ISI. Signal Over Noise.

• Revenue reached $41.5bn, up 346% on a year ago, at an 84.6% gross margin. No memory maker has held a margin near that through an ordinary cycle.

• Data centre revenue passed $25bn, around 61% of the company. The Cloud Memory unit alone earned $13.8bn at an 83% gross margin, against $3.4bn at 58% a year ago.

• Sixteen strategic customer agreements, fourteen carrying close to $100bn of minimum-price revenue with floors and deposits. HBM is sold out for all of 2026.

• High-bandwidth memory consumes roughly three times the wafer area per gigabyte of a DDR5 chip. AI demand therefore subtracts commodity supply. It does not add to it.

• The contracts that de-risk two years also commit about $27bn of FY26 capex to that demand holding. Samsung and SK Hynix are spending to match.

Subscribe now

The claim underneath Micron's record quarter is simple. The memory cycle will not break this time. It is worth testing that claim, because the industry has made it before.

Twice in the last decade the same belief took hold. In 2019, after a cloud build-out that looked permanent, DRAM contract prices fell by roughly half. In early 2023, after data centre demand was meant to have set a floor, Micron's gross margin dropped below minus 30% and the company wrote down $1.8bn of inventory. Memory is a capital race. Every supplier expands into the same shortage, and the shortage turns into a glut.

So the question for an 84.6% gross margin and a $50bn guide is not whether the print impresses. It is whether anything structural has changed since the last two crashes. Three things have.

The first is the contract.

Micron signed sixteen strategic customer agreements this quarter. Fourteen carry close to $100bn of minimum-price revenue, with floors, ceilings and deposits attached. These are take-or-pay commitments running years out, and they cover about a fifth of Micron's DRAM volume and a third of its NAND. Spot pricing set the last two cycles. Contracted pricing sets a growing share of this one.

The second is physical.

High-bandwidth memory consumes roughly three times the wafer area per gigabyte of an ordinary DDR5 chip. When the three makers who supply 90% of the world's memory shift capacity into HBM, they do not add supply beside the commodity market. They subtract it. AI demand and commodity scarcity have become the same event.

The third is who is asking.

In 2019 and 2022 the suppliers chased the demand. This quarter the buyers came early and asked to be locked in. A hyperscaler that pre-commits two years of memory has stopped treating it as a spot purchase. It now treats memory as a strategic input it cannot afford to be short of.

None of this repeals the capital race.

Micron will spend about $27bn on capex this year, and Samsung and SK Hynix are spending to match. Every prior glut began with exactly this confidence and exactly this spending. The floor is real for the bits that are under contract. It is soft for everything else.

That split is the part the record headline hides.

The contracted bits go to the hyperscalers. The squeeze lands on everyone without a two-year agreement. The carmaker, the industrial OEM and the PC builder now pay 60% to 95% more in a quarter for memory they once bought on demand. The AI memory boom is exporting a shortage to every buyer who is not a cloud.

Micron's quarter is not the top of a cycle or the end of one.

It is the moment memory stopped being one market and became two.

For three years the AI industry has been proving that software can do knowledge work. Code, text, images, analysis. Prometheus is a bet on the part of the economy those models never reached. The headline number is the least interesting thing about it. The company has no public product, no named customer and a deliberately thin footprint. What it has is a constraint that almost nobody is naming, and a structure built to break it. The constraint is data.

Subscribe now

Start with what it is not

The instinct is to file Prometheus next to OpenAI, Anthropic and the other frontier labs. That is the wrong shelf.

Bezos has been explicit. Today's large language models are good at knowledge work, which is why they help with code and maths. Prometheus is aimed at something he calls a different level of complexity: the design and manufacture of physical things. Jet engines. Chips. Medical devices. Cars. Drug compounds. The company's own phrase is "AI for the physical economy."

The product has a name: an "artificial general engineer." Not an assistant that writes about engineering. A system that does engineering. It takes a specification and the physics around it. It proposes designs, simulates them, optimises for weight, cost and safety, and pushes toward something a factory can build.

The prize Bezos names is time. The road from concept to mass production is long, and he wants to compress it tenfold. His example: ask a jet-engine maker for the same engine with ten per cent more thrust, and you may wait a decade. Prometheus wants that loop in months.

Why AI stops at the factory floor

Software ate the world. Now AI is eating it, but it stalled at the factory floor. The distinction matters, because it is the whole investment case.

The factory floor already runs on software: CAD, PLM, MES, the whole stack. What it never produced is the kind of tacit data a model can learn from.

Knowledge work moved its reasoning into text, so the models had a corpus.

The physical economy kept its reasoning in test rigs, in proprietary archives, in the heads of engineers. How a turbine is designed, why a casting fails, how a line is tuned, none of it was written down in a form an AI can read.

That is why manufacturing was never as easy a target as text, code or images. The software was already there. The training data was not.

This is the cleanest way to read Prometheus's bet. A chatbot learns from a near-infinite corpus. An artificial general engineer has no equivalent public corpus.

Bezos says the quiet part out loud: the company has to create its own data sets, and reach ones that are very hard to reach. Axios reports the same point in the company's own words, that there is no "internet of manufacturing data" to ingest.

The architecture underneath is deliberately vague, and Bezos is as clear about what it is not as what it is. When observers guessed Prometheus was building "world models," he said that is not right. He is firm that it has nothing to do with robotics.

Two things he will say. The system has to manufacture its own data. And internal benchmarks, he claims, already run some simulations faster than traditional methods, which is the line between a research idea and a tool an engineer will pay for. The labs most often named alongside it, Nvidia, Fei-Fei Li's World Labs, Covariant, Physical Intelligence, are working the same physical-AI frontier from different angles.

One concrete clue arrived by acquisition. In 2025, days after a single dinner, Prometheus bought General Agents, whose product, Ace, could operate a computer through video-language-action models. The capability is the tell. The thing that lets an agent drive a screen could let it drive the software an engineer lives in: CAD, a simulation suite, a product-lifecycle system. That is the bridge from a model that knows engineering to one that can do it inside the tools engineers already use.

The talent backs the ambition. Around 150 people across San Francisco, London and Zurich, with reported hires from OpenAI, DeepMind, Meta, Nvidia, Tesla and xAI. Two of the Transformer's authors, Ashish Vaswani and Jakob Uszkoreit, advise. Bajaj himself is a chemist and physicist who co-founded Verily, which tells you the lab-and-data angle is not incidental.

How it earns, and why that is the wrong question first

A $41bn valuation with no disclosed revenue is a claim about a business model, however coy the founders are. Three layers of earning are visible, and they get more interesting as they go. The whole model on one page:

Layer one: sell the tools. "The engineer is our customer," Bezos told CNBC. The first business is enterprise software for design, simulation and optimisation. The scope is the point. A jet-engine programme can run thousands of engineers for most of a decade. Shave a fifth off that bill and the saving is in the hundreds of millions, before a single gain in performance. On that maths there is room to charge in the millions a year per customer, though Prometheus has named no prices.

Layer two: become the engineering stack. Bezos likens the product to a very advanced form of CAD. That is a bigger claim than it sounds, because it points straight at Siemens, Dassault Systèmes, PTC, Autodesk and Synopsys, who own the installed base of design software. One route starts on top of the tools customers already run, proves value inside their workflows, and expands from there. The General Agents capability is what makes that route plausible.

Layer three, and the real one: own what it disrupts. This is where the coverage stops short. Alongside the lab, Bezos is reported to be raising a separate vehicle of up to $100bn to buy the industrial companies AI is about to reshape. The UAE's sovereign fund and JPMorgan's Jamie Dimon are named in the talks. Treat the fund as reported, not confirmed; Bezos declined to discuss it. He confirmed only a possible direction, while playing it down: the company may buy parts of businesses its technology can improve, and the plan is still in work.

Read those layers and you understand how Prometheus could earn. They miss why it would buy.

M&A is the crawler

The data thesis gives a second reason to own a factory, beyond the margin. It may be the bigger one.

An artificial general engineer needs proprietary industrial data, and that data does not live on the internet.

Language models scraped a finished web at near-zero cost.

Prometheus has no manufacturing internet to crawl.

If the reported holding company closes, the acquisitions could become part of the training set. In Henning Steier's phrase, M&A becomes the crawler and capital does the scraping. The Financial Times reported that the vehicle would have access to its portfolio companies' data to train Prometheus's models. Read that way, the reported structure is not only an investment vehicle. It is also a possible data pipeline.

Owning the factory would then pay for itself three ways at once.

(01) It gives the model rare data it cannot easily get elsewhere.

(02) It gives Prometheus a captive site to deploy the tools and prove them.

(03) And it captures the efficiency gains as the owner, where a vendor would only ever see a licence fee.

The data is the reason this is more than a buyout. The margin is what could make a very expensive data-acquisition pay for itself.

This is also the strongest case for the number. A pure software business is valued on software multiples. A business that owns rare data that could improve its product, and the assets that throw it off, is harder to copy, because the moat is not the model. Models get cloned and distilled. The data from running a real turbine is contractual, messy and scarce, and ownership is the cleanest route to control it.

Strip away the AI language and the shape is familiar: a Berkshire for the age of AI. Berkshire buys companies for their cash flows. This vehicle would be read as buying them for their data as much as their cash flows.

The bear case is not subtle

None of this is built. The honest version is that investors have paid $41bn for a team and a thesis.

The physical world has a clock speed, and it is slow.

The data is the whole game, and it is the hardest part to get. It sits behind NDAs, in formats no crawler reaches. Buying the companies that hold it is slow, capital-hungry and operationally brutal. Bezos calls it a capital-intensive start-up, which from a man who spent a decade funding rockets is saying something.

The target sectors are conservative by design. Aerospace and medical devices run on certification regimes that will not pass an AI-generated design because a model is confident. A hallucinated sentence is a typo. A hallucinated turbine blade is a fatality, and the regulator knows it. Adoption there is measured in years.

The structure carries its own risk. Running a frontier lab and a private-equity-scale acquisition machine at once is two companies, and Bezos is already splitting his time with Blue Origin and AI at Amazon. He says Prometheus is independent, with compute bought at arm's length, and has floated Blue Origin only as a possible case study customer. The conflict reading will follow a Bezos-owned company that wants to design and own factories, whatever the governance. Owning the rivals of your own customers is also an antitrust question, in industries that are already closely watched.

And there is the politics. Bezos told CNBC the productivity gains could turn two-earner households into one-earner households. He meant abundance. It will be read as a forecast of job losses. A company this size, proposing to automate engineering and own the factories, will carry that reading whether it wants to or not.

What would have to be true

First, the buyout vehicle. If the ~$100bn fund closes, the data-pipeline thesis becomes far more credible and the valuation starts to make sense. If it stalls, Prometheus is a very expensive bet on synthetic data and goodwill.

Second, a named acquisition or design partner. There are none yet. The first factory it buys or partners with tells you which data set it wants first. Watch chips, defence and aerospace.

Third, proof the loop compresses. One credible case where dream-to-build drops from years to months is worth more than another funding round.

Bezos built Amazon into a trillion-dollar company and bankrolled Blue Origin for two decades by being early, patient and willing to lose money for a long time. Prometheus is the same patience turned on the physical economy. In AI, capital is now the easy part. Data is the hard part. Prometheus is funded to create it, and the reported vehicle would let it buy its way to more.

Sources: CNBC (11 June 2026 interview and live coverage; earlier 20 May 2026 interview), cross-checked against Axios, GeekWire, TechCrunch, WIRED, Morning Brew and the Guardian. The reported ~$100bn buyout vehicle (WSJ, FT, Bloomberg, NYT) is unconfirmed by Bezos; the FT reported the data-access point. Figures from NYT, FT, WSJ, Bloomberg, Axios, GeekWire, TechCrunch.

The three things this period

1. SpaceX floated in a record IPO and is worth about $2.5tn, above Tesla. No multiple explains the price; it is a bet on the future.

2. Bezos's Prometheus raised $12bn to engineer the physical economy, and reportedly a $100bn fund to buy the data to train it.

3. The robots and the power got funded too. Neura took up to $1.4bn, KKR's Helix $10bn, Helion's fusion round $465m.

Subscribe now

Wall Street just priced a company on a different planet

On 11 June, SpaceX priced its IPO at $135 a share and floated on the Nasdaq. It raised about $75bn, up to $86bn if the bankers exercise their option, the largest IPO on record by proceeds. The shares jumped about 19% on day one, ran to an all-time high of $225 within days, and the company is now worth around $2.5tn, against Tesla's $1.49tn.

No multiple based on fundamentals explains it. The richest software companies trade near twenty times revenue. Hardware trades near two. SpaceX, at about 134 times, is on a different planet. SpaceX's launch business, the part you can watch, does not earn $2.5tn. The valuation is Starlink's recurring, subscription-like revenue, and a TAM that runs to Mars and to whatever AI demand puts into orbit. There is no public comp, because the market is not pricing a business. It is pricing a near-monopoly on the route to orbit, and a promise about what comes after. Call it a vote on the future.

So treat SpaceX as a special case, a once-in-a-generation asset valued on its own future, not a template for the sector. The repeatable signal this week is quieter, and it sits in the rounds nobody led with.

My read. SpaceX is the headline, not the lesson. Its price is a bet only it can underwrite. The money worth studying this week was priced on cash flows and ownership, not on a story about the 2030s.

Bezos is building a model for the physical economy, and buying the data to train it.

The same week, Prometheus raised $12bn at a $41bn valuation, past $18bn of capital since launch. The instinct is to file it next to OpenAI and Anthropic. That is the wrong shelf, even though the category is right. Prometheus is a frontier model too, aimed at a target those labs never touched. Bezos has been explicit: today's models are good at knowledge work; Prometheus is aimed at the part of the economy software never reached, the design and manufacture of physical things, jet engines, chips, medical devices, drug compounds. The product is an "artificial general engineer", a successor to CAD more than a chatbot. He is firm that it has nothing to do with robotics.

So in the first instance it is software. The bet that makes the $41bn legible is the data. An artificial general engineer needs proprietary industrial data, and that data does not live on the internet. There is no manufacturing corpus to crawl. Alongside the lab, Bezos is reported to be raising a separate vehicle of up to $100bn to buy the industrial companies the technology is meant to reshape, with access to their data to train the model. Treat that as reported, not confirmed. Read that way, the acquisitions are the training set, and the buyout is the route to data that does not yet exist.

Strip away the AI language and the question is an old one: where is the scarce asset?

With language models the data was effectively free, scraped from a finished web, so the model was the moat, and that moat erodes as models commoditise.

Prometheus starts from the opposite point. No corpus exists, so the scarce asset is whoever can assemble the proprietary industrial data, and buying the companies that hold it is the mechanism.

If the SpaceX price is a bet on a story. This is a bet on owning a bottleneck, the one data set a model of the physical economy cannot otherwise get.

My read. The exposed incumbents are the design-software owners, Siemens, Dassault Systèmes, Ansys and Autodesk, whose installed base the artificial general engineer is aimed straight at. The deeper move is the data. Owning the factory gives the model the proprietary data it cannot scrape, a captive site to prove the tools, and the efficiency as owner. M&A is the crawler. The margin is what makes an expensive data grab pay for itself.

If you want to go deeper, I will break down Prometheus business model in today’s article.

The deployment layer got funded too

Three robotics rounds closed inside three days. Neura Robotics, the German humanoid maker, raised a Series C worth up to $1.4bn on 10 June at about a $7bn valuation, with Tether, Qualcomm, Amazon and Nvidia in, alongside Bosch, Schaeffler and the European Investment Bank.

Bosch and Schaeffler are not financial investors. They are tier-one suppliers to every carmaker in Europe, and a cheque to Neura buys a position they would otherwise spend years building. A day earlier, Standard Bots raised $200m at a $1bn valuation for robots that do ordinary factory work. On 11 June, Theker added an $85m Series A for configurable factory robots, with Inditex and Samsung backing it.

My read. Watch the cap tables. A demo robot in a video is a sales asset. A maintained fleet is the business. The contest sits in the gap between the first convincing pilot and the thousandth boring deployment, where an automation strategy turns into operating leverage or dies in the integration queue.

And someone has to power it

All of it shares one constraint, and it is not chips. It is power.

On 10 June, KKR launched Helix Digital Infrastructure with more than $10bn of long-duration capital, alongside the Kuwait Investment Authority, Nvidia and Vistra. It is run by the former AWS chief Adam Selipsky, and it bundles data centres, power, transmission and fibre into one thing.

The scarce asset is now that bundle: a site, a grid connection, generation, cooling, permits, and a balance sheet large enough to carry it. Set it against the demand. Hyperscalers plan to spend close to $700bn on data centres in 2026, Amazon alone near $200bn. Helion, the Altman-backed fusion company, raised a $465m Series G on 4 June at a $15.5bn valuation to fund a plant for Microsoft. Last winter the US Department of Energy put $400m each behind small modular reactor projects from the Tennessee Valley Authority and Holtec.

My read. Energy is the layer to watch over the next year, because it gates the others. The route to software margins now runs through concrete, copper, steel, substations and heat rejection.

The backdrop: the capital is broad

Two more rounds frame the moment. Anduril raised $5bn in May at a $61bn valuation. Mach Industries raised $300m on 1 June at $1.8bn. Rockets, defence autonomy, factory robots, design AI, data centres and power are separate markets. They now compete for the same inputs: advanced chips, engineers, machine shops, actuators, power electronics and contract-manufacturing capacity.

My read. Five markets, one set of inputs. The same chips, engineers and machine shops now price against each other, and that squeeze is the quiet story under a loud week.

The prediction. Now that SpaceX has a public currency, watch for a move on Tesla. The IPO gave Musk a listed share to spend, he maxed his Tesla stake the same week. The window opens as the lock-ups expire...

^{Sources. SpaceX IPO: ISI SpaceX research. Prometheus: ISI deep dive research. KKR Helix: BusinessWire. Neura: CNBC. Standard Bots: PR Newswire. Helion: TechCrunch. Anduril: TechCrunch. DOE SMR (Dec 2025): World Nuclear News.}

The 20 May 2026 filing is the first audited disclosure of SpaceX's segment economics. Three segments, recast.

Connectivity (Starlink plus Starshield) generated $11.39 billion of revenue at 49.8% growth, $7.17 billion of segment EBITDA, 63% margin.

Space (Falcon, Dragon, Starship) generated $4.09 billion of revenue and lost $657 million at the operating line.

AI (xAI, Grok, X advertising, Colossus) generated $3.20 billion of revenue, lost $6.36 billion at the operating line, and absorbed $12.73 billion of capex.

Connectivity's $7.17 billion of EBITDA covers Space's losses with room. The same engine cannot cover AI's $6.4 billion operating loss plus $12.7 billion of capex. The IPO is the financing bridge to cover the gap.

SpaceX sits in a small set of multi-segment controlled-company structures at upper industrial scale, but the configuration itself is not new.

What is distinctive is the specific composition, the founder-controlled qualifier, and the cash position at IPO.

The cash engine itself exists because of an earlier strategic build nobody was funding at the time: reusable Falcon. The filing arrives just as the next bridge becomes a public-market problem.

The case the filing makes

The cross-subsidy operates in two layers. Connectivity carries Space; the operating loss in Space is roughly one-tenth of Connectivity's EBITDA. Connectivity cannot carry AI; the AI segment alone consumes more cash on operating account than the Connectivity engine throws off, before AI's capex line is added.

This is the structural read. The IPO is the mechanism that closes the gap Connectivity cannot. Use-of-proceeds in the filing ranks AI compute infrastructure first.

The pattern generalises even when the entity does not. Identify the cash engine. Decide what the engine is allowed to carry. Build a financing bridge for what the engine cannot cover. The SpaceX S-1 is the first audited disclosure of the company's own segment shape, and the cross-subsidy reads cleanly in the recast numbers.

What this shares with other multi-segment structures

The shape is not without precedent. Alphabet's other bets sit alongside the Search cash engine; Waymo, Verily, and the rest are financed against advertising revenue. Meta's Reality Labs carries a $17.7 billion FY24 operating loss financed by the Family of Apps. Amazon's retail expansion was financed for years by AWS segment cash.

The shape repeats outside US tech. Reliance funded Jio at multi-billion-dollar scale through external rounds against the oil and petchem engine. Berkshire's multi-segment portfolio has been audited under Buffett's single-controlling-shareholder structure for six decades.

The configuration is not new at scale either. Aramco priced at $1.7 trillion in 2019 under sovereign control. Alphabet's dual-class structure carries a $2 trillion market cap. Meta's founder-controlled structure carries $1.5 trillion. Reliance and Samsung sit at multi-hundred-billion-dollar caps under family control. Multi-segment plus controlled-company at upper industrial scale is a known shape.

The launch fleet is an internal Starlink utility

123 of 165 Falcon flights in 2025 carried Starlink for SpaceX's own account. Only 42 served external customers. Fleet reuse held above 85%. SpaceX's share of global mass-to-orbit was above 80%.

Starlink subscriptions cover the launch programme's fixed costs. External customers buy at marginal cost into a tempo SpaceX would be flying anyway. ULA, Arianespace, and Blue Origin have no commercial-broadband book of business funding their backbone. Until they do, the price floor for Western government and commercial launch is set inside SpaceX.

The strategy for challengers is hard. The price problem is structural. Matching SpaceX pricing requires either matching its scale (capital-intensive, ten-year horizon) or finding an equivalent internal subsidy line. No such subsidy line is visible in the public filings of ULA, Arianespace, or Blue Origin. The third option is a programmatic policy decision to underwrite an alternative provider. Any Western-capital conversation about Kuiper, Eutelsat, and the European launchers now starts from a hard fact. The price floor those alternatives have to beat is set by a cash engine the alternatives do not have.

The dependency runs deeper than the current cost allocation suggests. Starlink is viable as a cash engine because reusable Falcon dropped launch cost per kilogram by an order of magnitude over the decade before the constellation made commercial sense.

The launch programme paid for itself by enabling the cash engine that now subsidises it. The sequencing is the pattern: strategic build first, cash engine second, current cross-subsidy third.

The AI capex going in today is enabling infrastructure for a the new (deemed) growth engine.

Starlink is the engine

10.3 million subscribers across 164 countries by 31 March 2026, doubled from 15 months earlier. Consumer ARPU compressed from $99 in 2023 to $66 in Q1 2026, a 33% drop. The most recent quarter was the steepest single step, down 19% from the full-year 2025 number.

The 63% segment EBITDA margin survives the consumer ARPU compression because the enterprise tier carries the margin. Maritime, aviation, and direct-to-device are higher-priced, higher-margin lines. Direct-to-device hit 7.4 million monthly unique devices across 30 countries in Q1 2026. The growth story is volume against a falling consumer price. Enterprise absorbs the margin work.

The lesson for businesses studying the platform shape is clear. ARPU pressure on the consumer flank is survivable when two conditions hold:

The enterprise flank carries the margin.

The volume on the consumer flank keeps the network economics working.

A consumer-tier price war on its own would compress margin. The filing shows it can be absorbed when those two conditions are met.

The federal share is one-fifth

The S-1 frames federal revenue as approximately one-fifth of 2025 revenue. That implies $3.7 billion against $18.7 billion consolidated. Cumulative life-of-programme awards run $24.4 billion across NASA Commercial Crew and Cargo, Artemis HLS, NSSL Phase 3 Lanes 1 and 2, NRO Starshield, SDA Tranche, and PLEO. NSSL Phase 3 Lane 2 added $5.9 billion in April 2025. Golden Dome AMTI added $2 billion in July 2025.

The framing matters because the common read of SpaceX still treats it as a government contractor with a side commercial business. The filing inverts that. 20% of revenue says SpaceX is a commercial business with a federal tail. Western governments considering how to underwrite alternative providers should price in a smaller demand-side lever than they used to hold. The Pentagon is one customer among many.

AI is the capital ask connectivity cannot cover

X advertising contributed $1.84 billion of the AI segment's revenue. The Anthropic Colossus contract contributed the bulk of AI Solutions and Infrastructure's $1.36 billion.

Anthropic pays $1.25 billion per month through May 2029, terminable on 90 days notice either side. Headline run-rate is $15 billion per year.

The 90-day termination clause runs both ways. SpaceX rented compute capacity it will pull back as soon as xAI needs it. Anthropic is buying compute on a lease its supplier can reclaim. Both counterparties carry the contract risk.

Net income flipped from $791 million of profit in 2024 to a $4.94 billion loss in 2025. Cash on hand fell from $24.7 billion at the end of 2025 to $15.9 billion by 31 March 2026, an $8.8 billion drop in one quarter. Operating cash flow in Q1 2026 was a positive $1.0 billion. Investing cash flow in Q1 2026 was a negative $16.7 billion. The burn sits on the capex line; the operating line is in surplus. The Connectivity engine cannot cover the programme on operating account alone. The IPO closes the gap.

The capex line is what the IPO funds. $12.73 billion in 2025. $7.72 billion in Q1 2026, annualising above $30 billion. Servers and networking on the balance sheet at $23.85 billion. Construction in progress at $14.05 billion. Contractual commitments at $25.45 billion against a $15.9 billion cash balance.

The financing pace is part of the strategic read. $28 billion in 2018. $800 billion in December 2025. xAI merger close at $1.25 trillion in February 2026. S-1 indicative range $1.75 trillion to $2.0 trillion at filing. The valuation has roughly doubled in five months. The build the IPO funds is sized to that pace.

For boards observing what a multi-segment strategic build financed publicly looks like at upper scale, the SpaceX shape is a precedent worth studying. The cash engine has to be generative enough that the financing bridge is plausible. When the engine cannot bridge, the strategic build dilutes or the company is forced to public markets earlier than planned. SpaceX has been forced earlier than the timeline it could have run privately.

The largest founder-controlled IPO ever priced

Musk holds 12.3% of Class A and 93.6% of Class B. Class B carries 10 votes per share. Aggregate voting power is 85.1% on a 42 to 43% economic stake. Class B holders elect a majority of the board. SpaceX will claim controlled-company status under Nasdaq rules. Mandatory arbitration applies. No class actions. No jury trials. No planned dividend.

Musk's lock-up is 366 days with no early release. Standard insider lock-up is 180 days with tiered early release up to 65% after Q3 2026 earnings. Musk's 2025 base salary was $54,000. The January 2026 performance grant is one billion Class B shares with a Mars colony of one million inhabitants as the headline vesting condition.

Dual-class structures at IPO are not new. Meta, Snap, Palantir, and Coinbase all listed under variations of the model. Aramco priced larger than SpaceX in 2019 under sovereign control. Among founder-controlled IPOs, SpaceX raises the structure to $1.75 trillion at first trade and becomes the upper-end precedent for the format. Boards considering how the next decade of founder-controlled public-market governance is structured should know this is now the precedent at the top of the range.

The Cursor (Anysphere) acquisition option sits on the cap table separately, carrying $60 billion of potential consideration with $10 billion of walk-away exposure. It is a contingent commitment in the disclosure pack, not in the audited segment economics. Reading the filing as a precedent means reading the contingent stack alongside the recurring stack.

The pattern worth studying

What generalises is the pattern, not the entity.

Profitable segments finance future strategic builds. Then, external financing is sought when that ceiling is reached. The public market (IPO) is the financing instrument here.

These questions belong on the next board agenda for any industrial firm looking to make future bets:

What is the cash engine in the portfolio?

What is the test for the engine being asked to carry too much?

And which strategic build today is laying the rails for tomorrow's cash engine?

Subscribe now

KEY INSIGHTS

The flip: Seven of the top 10 reported positions are now puts on the AI compute stack. SMH ETF, Nvidia, Oracle, Broadcom, AMD lead the short complex. Put notional totals $8.46bn; long equity totals $3.86bn.

The rotation: Long money concentrates in bitcoin miners pivoting to AI hosting, in Bloom Energy fuel cells, in CoreWeave, in US solar via T1 Energy, and in mobile gas turbines via Solaris Energy Infrastructure. CleanSpark up 648%, Bitfarms up 188%, Riot up 86% versus Q4 25 share counts.

The Intel flip: The single sharpest signal. Long $747m of Intel calls at Q4 25 close. Short $159m of Intel puts at Q1 26 close. Same name, opposite direction, one filing apart.

The fund-size caveat: The headline $13.7bn reported value is not deployed capital. ADV-filed AUM is roughly $383m. Most of the apparent Q1 26 growth is put-notional overlay. The long equity book actually shrank slightly versus Q4 25

Subscribe now

The brief

The Situational Awareness essay (June 2024) named the binding constraint on the AI buildout. Compute would gate everything. The gatekeepers of compute would be the most-bid names of the decade. Power would be the operational headache that made the buildout slow. The fund that took the essay’s title has just put $8.46bn of put notional on the gatekeepers and concentrated the long book in the operational layer. Read as an industrial-strategy signal, the shape of the book is the thesis updating in public.

Walk the top 10.

SMH put at $2.04bn. Nvidia put at $1.57bn. Oracle put at $1.07bn. Broadcom put at $1.01bn. AMD put at $969m. Bloom Energy long at $879m. SanDisk long at $724m. Micron put at $584m. CoreWeave long at $556m. TSMC put at $535m.

The visual asymmetry of seven puts to three longs in the top 10 is the industrial claim, in capital form. Concentration sits very high. Top 10 lines equal 72.7% of reported value. Top 5 lines equal 48.7% and are all puts on the chipmaker complex.

The new options overlay is the structural news of the quarter. The $8.46bn of put notional did not exist at Q4 25. The $1.36bn call layer on Micron, SanDisk, and TSMC also is new. In one filing the fund added an options book equivalent in notional terms to more than twice its long equity. Quiver Quantitative led with the size of the bet. TrendSpider catalogued every put line. Bloomberg’s news desk read it as a directional short. The Motley Fool and Anna Coulling read the same data as a hedge.

The data alone does not settle that argument.

A 13F-HR reports the fair value of long equity and the notional value of options. It does not disclose whether the manager is long or short the option line. Brandon Gorrell at TBPN, TrendSpider, and several others flagged the caveat on filing day. The trade-structure question matters less for you as an industrial-strategy reader, than the long-book rotation does.

Three observations still weight the directional-short reading over the pure-hedge reading. The puts cluster on the chipmaker stack. A hedge against a long AI book would distribute exposure more evenly. The Intel single-name flip from long call to short put has no clean hedge interpretation. And the long book has been rotated over two quarters into the physical infrastructure layer that runs the chips. The hedge reading does not explain the rotation. The thesis-update reading explains both.

Map the holdings onto the bottlenecks named in the original essay.

Compute, power, algorithmic efficiency, security. The compute bucket is $8.45bn short via puts against $785m long. The power, data centre, and miner-pivot buckets are $3.07bn long against zero short. Each long sits on a physical-capacity claim.

Bloom Energy at $879m is solid-oxide fuel cells delivering power behind the meter in roughly 90 days against utility interconnect queues of 2-5 years. I wrote about this on Monday ↓

Solaris Energy Infrastructure owns mobile gas turbines rented to data centres for emergency and interim load. T1 Energy is US solar PV manufacturing. Babcock & Wilcox sits adjacent to small modular reactor work. CoreWeave is AI-native cloud capacity. The bitcoin-miner-pivot cluster owns pre-existing grid connections and gas-fired or behind-the-meter generation that can be repurposed to AI hosting without the permit cycle.

The portfolio shape tracks the industrial reality of the buildout.

Power transformer lead times in 2026 sit at 128 weeks. Generator step-up transformers at 144 weeks. Switchgear at 44 weeks. The Sightline Climate data-centre pipeline shows 12 GW of capacity announced across 140 US projects with only 5 GW actually under construction. Three of the seven announced Stargate sites are running on on-site natural gas microgrids precisely to bypass grid interconnect queues. The Crusoe and Engine No. 1 and Chevron joint venture supplies 4.5 GW of gas to Stargate Abilene through seven GE Vernova LM2500XPRESS aeroderivative turbines. Bloom’s October 2025 partnership with Brookfield committed $5bn of fuel-cell deployment across AI data centres. Oracle’s Project Jupiter order ran up to 2.85 GW of Bloom capacity. The book is long the names that solve the operational problem the essay named in 2024.

Two qualifiers sit alongside the picture for any reader treating this 13F as an industrial-strategy signal.

The first is fund size. Mark Pacitti’s LinkedIn write-up flagged the gap that several other observers have since echoed. The fund’s ADV-filed AUM is $383m. The Q1 26 long book is $3.86bn. The gross-exposure-to-AUM ratio implies leverage in the order of 14 times. The headline “$13.7bn fund” framing now widespread in finance press materially overstates the deployed capital base.

The second qualifier is the inference-scaling complication that the original essay did not anticipate. Toby Ord’s March 2025 work and the Epoch AI analyses argue that capability gains since GPT-4 have come predominantly from inference-time compute. Pre-training scaling has slowed.

If that reading is right, the era of the trillion-dollar single cluster may already be past peak, the power demand profile reshapes, and the binding constraint in the system shifts again. The portfolio is positioned for the original physical-capacity story. The inference-distributed scenario sits outside the bet.

The bottleneck framing predates the Situational Awareness essay by forty years.

Eliyahu Goldratt named it cleanly in The Goal (1984) as the Theory of Constraints.

Every system that produces something is, at any given moment, limited by one binding constraint. Throughput is set by the binding link. Lift it, the constraint moves to the next weakest.

Goldratt’s five-step cycle: identify the constraint, exploit it, subordinate everything else to it, elevate it, repeat at the new binding link.

The 13F is the cycle running in capital-market form. Aschenbrenner identified the binding link in June 2024. The capital response was so violent that the constraint moved. The Q1 26 portfolio is the second-cycle move. Long the new binding link. Short the no-longer-binding link, because the price of the latter still embeds a binding-constraint premium that no longer applies.

The bottleneck framing shows up across the commentary on this filing without being named.

Goldman Sachs Alternatives wrote a memorable line on the AI buildout in 2026. Chip designers, memory manufacturers, and fabs capture roughly 90% of the profit pools and represent 10% of the chokepoints. Power, grid, and mission-critical services capture 10% of the earnings and 100% of the chokepoints. That is Goldratt language without the attribution. Anna Coulling described the fund as supplying the picks, shovels, and the power plants while hedging the cost of the gold itself.

Jordi Visser at Visser Labs wrote that the next decade will not be defined only by who writes the best code. It will be defined by who can build, power, cool, connect, manufacture, and deploy the physical infrastructure.

Theory of Constraints is one of the lenses I use often when advising industrial operators. It fits this filing cleanly because Aschenbrenner has run the cycle in public.

Three checks you can implement today in your industrial business, after reading this filing.

(01) First, identify where throughput is queueing this quarter. (02) Second, name which adjacent capacity becomes the new bottleneck when the current one is lifted. (03) Third, audit whether the operation is elevating the binding link or optimising a link that already has slack.

I work a version of this check with most of the advisory clients I sit with. The same logic runs on a factory floor as in a 13F.

If you found this article useful, please share it.
And you have subscribed to ISI already, you can do so for free on the link below ↓

Subscribe now

A NOTE FROM DAN

This edition consolidates two weeks of material, as I have been focused on client fundraising activities.

If you’re an industrial or deep-tech team looking to raise capital in 2026, I will be launching this as an offering - INDUSTRIAL RAISE™ - to work with you on

(01) Position → (02) Raise → (03) Scale

THE FORTNIGHT SUMMARISED

Volkswagen’s CFO told the market its announced cost cuts are “not enough” as Q1 profit fell 14%, hours before Donald Trump moved EU auto tariffs from 15% to 25%. ArcelorMittal called the steel quarter a “trough” the same week the EU agreed a 50% out-of-quota duty that will take 13mt of imports off the continent - whilst BYD opened talks to absorb a shuttered VW plant in Dresden. Germany’s Ifo and ZEW indicators both printed at multi-year lows three days apart. OpenAI paused its UK Stargate flagship over industrial power costs, and Microsoft absorbed the displaced Norway capacity inside five days. The Iberian grid’s one-year anniversary verdict landed with no accountability as Brent intraday touched $126 on Hormuz.

Many of the European industrial pillars took a hit in the same fortnight. The continent’s room to manoeuvre on cost, capacity, and capital is narrowing.

SIGNAL

The stories worth attention.

Caterpillar reported Q1 2026 sales of $17.4bn on 30 April, up 22% year-on-year, with adjusted EPS of $5.54. The number that matters sits in the order book. Backlog hit a record $62.7bn, up $11.5bn sequentially and $27.7bn year-on-year. Power & Energy alone was up 22% to $7.0bn, with power-generation revenue up 41% to $2.82bn. Six separate customer agreements have now been signed for at least 1GW each.

CEO Joe Creed used the print to upgrade the company’s reciprocating engine capacity plan from “double the 2024 base” to “triple it” by 2029. This is the second upgrade in 18 months. Capex is now earmarked at 4-5% of MP&E sales through 2030, and the 2030 enterprise revenue CAGR target was raised to 6-9%. Tariff drag on Power & Energy margin (down to 20.6% from 22.3%) was absorbed inside Construction Industries, where margins expanded to 21.4%.

The doubled-then-tripled capacity plan is the cleanest signal yet that hyperscaler-led behind-the-meter power is now a multi-year manufacturing capex commitment. It also reframes what Caterpillar is. The construction cycle is no longer the dominant capital allocation story; gensets for AI are. Order books are extending into 2028, and the binding variable on hyperscaler firm-power timelines is now Caterpillar’s factory throughput, not its dealer network.

The read-across is direct for Cummins, Wärtsilä, MTU and Rolls-Royce Power Systems, all exposed to the same demand and none yet committed to a tripled plan. For UK readers, the engines, not the silicon, are the long-lead item on domestic data centre buildout, and the supply runs through Lafayette and Griffin, not Reading.

My read: Caterpillar is temporarily becoming an AI infrastructure company with a construction division attached. The board signal is that hyperscaler-led demand for firm power is structural, large enough to justify two consecutive capacity upgrades inside 18 months, and tightly coupled to a customer set that signs 1GW-block agreements one at a time. UK and EU strategists who still treat data centre siting as a planning problem are working on the wrong constraint.

Caterpillar Reports First-Quarter 2026 Results

ACROSS THE SECTORS

NATO’s Support and Procurement Agency is in advanced selection for the replacement of its 14-aircraft Boeing E-3 Sentry AWACS fleet, with the Saab and Bombardier GlobalEye reported by French defence outlet La Lettre on 23 April as the chosen platform. Saab’s VP of media relations publicly stated on 24 April that no contract had been signed and that Saab had only “provided information”. Multiple trade outlets have repeated the selection report citing programme value above €5bn (USD 5.7bn) for 10 to 12 aircraft, with entry into service projected for the early 2030s. The NATO competition reopened after the US cancelled its own E-7 Wedgetail acquisition in June 2025.

What is firm is the direction of travel. NATO’s airborne surveillance backbone has been a Boeing programme since 1982. The fact that Saab/Bombardier is now the lead candidate, with no European industrial veto on a Canadian-Swedish prime, is the structural signal. Pair it with Germany’s leaked posture of routing only 8% of new equipment value through US suppliers and the message is that European procurement is now actively choosing alternatives where they exist.

For Boeing, this is the most visible defence prime loss of the year and the loudest signal that the political risk premium on US platforms inside NATO has gone up. Sustainment and training revenue for the Erieye Extended Range radar and Global 6000 airframe is long-cycle, into the 2050s.

NATO eyes Swedish-Canadian jet for AWACS role in shift away from Boeing

Rheinmetall signed a framework contract with the Bundeswehr in Koblenz on 22 April for FV-014 loitering munitions. Programme ceiling is described as “in the billions of euros”, with a first call-off of approximately €300m booked the same day. The contract optionally covers a five-figure quantity of FV-014 autonomous reconnaissance and strike drones with a 100km range and 70-minute endurance. On 27 April, Germany’s BAAINBw confirmed a €1.04bn (USD 1.18bn) call-off for the Infantry Soldier of the Future programme, covering 237 platoon kits with deliveries 2027 to 2029.

Defence Minister Boris Pistorius presented Germany’s first standalone military strategy on 22 April, naming Russia as the “biggest and most immediate threat”, setting 2029 as the deadline for Bundeswehr large-scale conflict readiness, and a 2039 target to become Europe’s strongest conventional fighting force. Pistorius also flagged that surging Middle East demand for air defence is now compressing global production capacity, an unusual public admission that the binding variable for European industrial planning has migrated from orders to throughput.

Two contracts above €1.3bn in a single fortnight, against a doctrine fixing 2029 as a war-readiness deadline, is the clearest evidence yet that German defence demand is moving from announcement phase to revenue phase. For European primes, the framework agreements signed in 2024 and 2025 are now generating large quarterly call-offs that will lift order books through 2027.

Rheinmetall lands $1.2 billion German future soldier contract

Volkswagen Group reported Q1 2026 operating profit of €2.5bn ($2.92bn), down 14.3% year-on-year and well below consensus near €4bn. Sales revenue fell 2.5% to €75.7bn. CFO Arno Antlitz told the market that cost reductions already underway, including roughly 50,000 German job cuts by 2030, are “not enough” and that VW must “fundamentally change” its business model “or jeopardise our future”. US tariffs added €1.3bn in the quarter alone, with Audi, VW Passenger Cars and Cariad booking a further €0.7bn of restructuring provisions. China deliveries fell 15%. CEO Oliver Blume publicly endorsed sharing idle European plants with Chinese OEMs as a “clever solution”.

Stellantis the same day reported Q1 net revenue of €38.1bn (+6%) and a return to profit at €377m, but warned net 2026 tariff cost still totals €1.3bn, revised down only after a $0.4bn IEEPA refund tailwind. Combined VW-Stellantis 2026 tariff exposure now sits above $5.9bn. On 1 May, the US moved the EU auto tariff from 15% to 25%, announced by social media without implementing detail.

Two of Europe’s three largest carmakers conceded in public on the same day that announced restructuring plans are inadequate. Treat European OEM restructuring as a multi-year compounding exercise. The single rounds of cuts already announced will not close the gap. The Blume “share factories with Chinese rivals” line is the loudest signal yet that German management views Chinese tie-ups as a survival lever, with implications for supplier base, IP, and the political contract underpinning Made-in-Germany.

Volkswagen warns planned cost cuts are not enough after 14% drop in first-quarter profit

Carnewschina reported on 1 May that BYD is in early-stage talks with Volkswagen to take over part of the Transparent Factory in Dresden, where VW stopped car production in 2025. The proposal would split the site between TU Dresden as an innovation hub and BYD for EV assembly, with conversion costs around €50m under discussion. The story landed the day after Blume publicly endorsed factory-sharing with Chinese OEMs. VW has denied a separate Reuters report on a different plant tie-up, and the Dresden talks are sourced to unnamed parties.

The volume side of the same window is firmer. BYD reported April overseas passenger and pickup sales of 134,542 units (+70.9% year-on-year), now 42.8% of group volume, even as domestic-led NEV sales fell 15.5% for the eighth straight monthly decline.

BYD net profit per vehicle is estimated to have collapsed to roughly RMB 3,000-4,000 in Q1 2026 from RMB 8,800 a year earlier, with Q1 net profit down 55% to RMB 4.08bn ($597m).

A Chinese OEM operating an assembly footprint inside Germany, with a German OEM as landlord, sidesteps the EU’s 17% anti-subsidy duty plus 10% standard import tariff and converts BYD from importer into local manufacturer in the symbolic core of European automaking. Even at MoU stage, the optics reset the policy debate. For European tier-one suppliers, the de-globalisation of EVs is happening through factory absorption, not greenfield. Capital allocators in European auto real estate, suppliers, and tooling should expect more conversions.

BYD in talks to take over part of Volkswagen’s Dresden plant

AgiBot announced on 14 April that its G2 humanoid had completed an eight-hour live-streamed shift on Longcheer Technology’s tablet production line in Nanchang, performing precision loading and unloading at MMIT testing stations. The company said G2 has accumulated 140 hours of continuous operation alongside human operators, with deployment expansion to 100 units planned by Q3 2026. At the AgiBot Partner Conference on 17 April, founder Peng Zhihui declared 2026 as “deployment year one” for embodied AI productivity, with subsequent rollouts targeting automotive, semiconductors, and energy.

This is precision millimetre-level work inside a real EMS manufacturing flow, on a public live-stream, in the same fortnight Tesla pushed Optimus production back to late summer and Boston Dynamics’ Atlas remained pre-production for Hyundai. Unitree’s prospectus shows its application revenue still concentrated in tour-guide and reception use. AgiBot is staking the industrial deployment crown ahead of Unitree’s IPO.

The competitive pattern across the window is brain layer winning in the West and body layer winning in China. China is operationalising what Western pure-plays are still preparing for. UK and European industrial strategists should treat factory-floor humanoid deployment as a 2026 buyer’s market for vertically integrated Chinese vendors, not a 2027 problem.

Xinhua: Chinese humanoid robots deployed on assembly lines for precision tasks

Skild AI acquired Zebra Technologies’ robotics automation business on 15 April, including the Fetch Robotics AMR line that Zebra bought for $290m in 2021 and the Symmetry Fulfillment orchestration platform. Financial terms were not disclosed. The deal follows Skild’s $1.4bn Series C at a $14bn-plus valuation in January, backed by SoftBank, NVIDIA, Bezos, Lightspeed, Coatue, Sequoia, Samsung, LG, Schneider, and Salesforce Ventures.

Skild now has a single stack covering humanoids for pick-place, quadrupeds for inspection, robotic arms for packing, AMRs for material movement, and an orchestration layer that talks to wearable devices on human workers. This is the first deal in the humanoid era to bundle the foundation model, the multi-form robots, and the warehouse-scale orchestration software in one entity. It is also the architecture every operator says it actually wants. Logistics buyers, Walmart, GXO, DHL, and Maersk, can now sole-source a stack from one supplier and skip three-vendor integration.

Zebra exits hardware-led robotics after a five-year underperformance, freeing capital, while Skild moves from pure model vendor to platform. For UK and EU industrial software vendors looking at warehouse automation, the procurement question shifts from “best-of-breed integration” to “compete with a foundation-model-led stack”. Aveva, Cognite, ABB Ability, and Siemens Industrial Edge teams should be benchmarking their orchestration roadmaps against this transaction this week.

Skild AI Acquires Zebra Technologies’ Robotics Automation Business

The European Parliament and Council reached political agreement on 13 April on the new steel measure replacing the expiring safeguard from 1 July. Tariff-free quotas are set at 18.3 million tonnes per year across 30 product categories. The out-of-quota duty is 50%, with a “melt and pour” traceability requirement. ArcelorMittal’s Q1 2026 results on 30 April characterised the quarter as a “trough”, reporting net income of $0.6bn and EBITDA per tonne of $131. Management projected CBAM plus the new TRQ would cut EU steel imports by approximately 13mt versus 2025 levels. The company took a final investment decision on a new electric arc furnace in Dunkirk and flagged $1.7-1.9bn of strategic capex for FY26.

This lands eleven days after Donald Trump signed a 2 April proclamation restructuring US Section 232 tariffs (effective 6 April): 50% on articles made entirely or almost entirely of aluminium, steel, or copper; 25% on derivatives; 15% through end-2027 on metal-intensive industrial equipment; and 10% on products made abroad with American-origin metal. The April ISM Prices Index jumped 6.3 points to 84.6, its 19th straight monthly increase.

Both Western blocs erected metals walls in the same fortnight. The EU has effectively converted its steel market into a closed system: CBAM from 1 January 2026, a 50% out-of-quota duty, and 47% lower tariff-free quotas. Mill margins in Europe should re-rate. Dumping pressure pivots to India, Turkey, and ASEAN. Chinese exporters face a now-systemic 13mt hole. UK machine-tool buyers sourcing US-origin equipment should expect 2026-2027 list price increases.

Council and European Parliament strike deal to protect EU’s steel industry from global overcapacity

SK Hynix reported Q1 2026 revenue of KRW 52.6tn (around $36.7bn) on 23 April, up roughly 60% sequentially and 198% year-on-year, with operating margin of about 72%. DRAM ASPs rose by a mid-60% range sequentially. On the call, management disclosed that cumulative HBM purchase commitments from hyperscalers and accelerator vendors now exceed three years of nameable forward supply. HBM4E samples ship in 2H 2026 with mass production targeted for 2027.

Samsung Electronics reported on 30 April: revenue KRW 133.9tn (around $90.0bn), up 50.2% year-on-year, and operating profit KRW 57.23tn (around $38.5bn), up roughly eightfold. The Device Solutions semiconductor segment delivered operating profit of KRW 53.7tn, up about 49 times year-on-year. Samsung confirmed it is mass-producing HBM4 alongside HBM3E, supplying NVIDIA, Google, and AMD. Management noted that for now conventional DRAM is delivering higher per-unit margin than HBM, because HBM is locked into annual price cycles while DRAM reprices on shorter terms.

A multi-year HBM order book held by a single supplier removes pricing optionality from every AI accelerator buyer without second-source qualification. The seasonal Q1 dip did not arrive. This is the cleanest evidence yet that 2026 AI capex is structural, not cyclical, and that memory, not logic, is the binding constraint. Samsung’s HBM3E qualification plus HBM4 in production ends the SK Hynix monopoly on top-bin HBM and pulls every server, PC, and handset OEM into a tighter procurement environment.

CNBC: SK Hynix posts record first-quarter profit, in line with estimates as memory prices climb

(07) SEMICONDUCTORS & COMPUTE

EU Chips Act 2.0 will let Brussels take direct equity in fabs

Bloomberg reported on 30 April that the European Commission’s Chips Act II proposal, scheduled for tabling in late May, would for the first time give the Commission authority to invest directly in cross-border semiconductor manufacturing. Until now Brussels has only coordinated and approved state aid. The redesign also broadens scope to upstream technologies including semiconductor equipment, materials, and advanced packaging. Industry submissions to the consultation have proposed an EU-level envelope of €30-60bn plus member state contributions of €50-60bn over seven years.

The redesign follows the European Court of Auditors’ finding that the original 2022 Chips Act is unlikely to meet its target of doubling Europe’s share of global chip output, and after several flagship fab projects in Europe were delayed or cancelled. Direct Commission equity is a structural change in European industrial policy. It moves Brussels from gatekeeper to investor and shifts who carries fab construction risk.

For ASML, Infineon, STMicro, Soitec, and the Dresden ESMC project, the proposal sets the financial baseline for the next decade of European chip sovereignty arguments. UK readers should watch whether London is pulled into the framework or shut out. Pragmatic in South Wales, ARM, and the UK compound-semiconductor cluster sit on the edge of the policy depending on how cross-border eligibility is drafted. The EU has switched from coordinate-and-subsidise to ownership-and-restrict, the same posture the US adopted in 2024 with the BIS export control regime. The UK has not yet picked its lane.

EU Chips Act Revamp Would Let Commission Invest Directly in Fabs

(08) DATA CENTRES & DIGITAL INFRASTRUCTURE

OpenAI walks UK Stargate; Microsoft absorbs Norway in five days

OpenAI paused its UK Stargate site at Cobalt Park in the North East AI Growth Zone on 9 April, citing UK industrial electricity costs and unresolved AI copyright regulation. Five days later, Nscale and Microsoft confirmed Microsoft would take the 230MW Narvik (Kvandal) campus that had been earmarked for OpenAI’s Stargate Norway, with Microsoft contracting an additional 30,000 NVIDIA Rubin GPUs on top of the $6.2bn Nscale deal it signed in September 2025. The new Rubin tranche is scheduled for 2027 delivery. On 16 April, UK AI minister Kanishka Narayan rebutted OpenAI publicly at a Bloomberg event in London, saying nothing in the UK energy or regulatory environment had changed since OpenAI’s original commitment. OpenAI confirmed in the same week it would still open a permanent London office.

All four leading hyperscalers raised 2026 capex guidance on 29 April. Microsoft now around $190bn, with $25bn of the lift attributed to memory pricing alone. Alphabet $180-190bn, Meta $125-145bn, Amazon roughly $200bn. Combined: $650-725bn. Anthropic separately locked in 10 GW of training power and around $65bn of equity inside five days, between Amazon’s $25bn top-up on 20 April and Google’s $40bn announcement on 24 April.

The UK’s AI Growth Zone framework cannot deliver hyperscale AI compute at viable cost in its current form. Industrial power prices roughly 75% above pre-Ukraine baseline, plus a 12-15 year grid-connection queue, are now sufficient to drive frontier-lab capex elsewhere. Microsoft picking up the displaced Norway capacity within five days shows demand is unchanged; the substrate has shifted to whichever European jurisdiction can deliver firm cheap power fastest. UK ministerial response was political, not policy, which leaves the underlying arbitrage intact.

Forward Prediction: The Cobalt Park GPUs originally earmarked for OpenAI UK will be reallocated to Norway or Ireland by end-Q3 2026. If they are not, there is no UK industrial-strategy answer to the Stargate exit; the Wylfa SMR contract becomes the only firm-power lever the UK has left.

OpenAI halts UK Stargate project amid regulatory and energy price concerns

(09) ENERGY, NUCLEAR & GRID

Iberian blackout anniversary lands with no accountability and an ENTSO-E verdict that quietly indicts the grid

Spain and Portugal marked one year since the 28 April 2025 blackout that shed 31 GW in five seconds and cut power to 47 million people for up to 16 hours. ENTSO-E’s 49-member expert panel published its final report on 20 March 2026, blaming rapid voltage rise, gaps in reactive power and voltage control, generator mismatches, and cascading disconnections in Spain that the existing defence architecture could not arrest. The report explicitly cleared renewable penetration as a root cause. As of the anniversary, none of the 30-plus sanction proceedings opened by Spain’s CNMC against grid operator REE and ten energy companies has produced a final ruling. Portugal has committed €400m to grid investment, including expanding battery storage to 750 MW. The Bay of Biscay HVDC interconnector to France, sized at 5 GW after a June 2025 capacity uplift, remains on track for 2028 commissioning.

The week ran in parallel to a major oil and gas backdrop. Brent rose for eight consecutive sessions through the window as US-Iran negotiations stalled and the Strait of Hormuz remained effectively shut. On 30 April Brent intraday touched $126.41, the highest since 2022, before settling 3% lower at $114.01. The IEA called the disruption the largest supply disruption in history; global oil supply fell 10.1 mb/d in March. The UAE’s separate decision to leave OPEC during the window added a structural overhang.

For UK and EU industrial buyers, the fortnight is the cleanest evidence yet that grid resilience risk in 2026-28 sits with voltage stability and inverter behaviour, not generation mix. Siting decisions across Iberia, Ireland, and the wider European fringe should now price a credible blackout premium until the Biscay link energises. Reactive-power compliance is the next regulatory front, against a backdrop of 128-week transformer order books.

Anatomy of a Blackout: Findings from the Spain-Portugal Grid Collapse Final Report

THE WEEK AHEAD

EARNINGS TO WATCH

MACRO

• US ISM Services PMI, Monday 5 May. Pair with the manufacturing print at 52.7 to gauge whether the services-side weakness flagged in the April composite is broadening.

• Bank of England decision, Thursday 8 May. Markets price in a hold; the inflation note will set tone for industrial financing costs.

• Eurozone retail sales and German industrial production, Friday 9 May. The first hard-data confirmation of whether the Hormuz-driven April PMI strength is real or stockpiling.

EVENTS

• Hannover-Stuttgart-Munich industrial trade-show season continues. Watch for follow-through customer announcements on agentic-manufacturing architectures presented at Hannover Messe (20-24 April).

• EU AI Act trilogue resumes, Wednesday 13 May (just outside the week). Industrial AI vendors and operators planning against an assumed deferral need a second-trilogue decision before May board reviews.

ON THE RADAR

• BYD Q1 2026 dividend and capex guidance update following the Dresden talks and 55% Q1 net profit decline. Capital allocation between domestic discounting and overseas expansion is the question.

• US Commerce Department response to the Senate Section 232 letter on CNH, Caterpillar, and Deere Mexico production. Any preliminary determination would reset 2026 list prices for North American agriculture and construction equipment.

• ACEA EU passenger car registration data for April, expected mid-May. First confirmation of Tesla’s continued leadership loss to BYD in select European markets.

• OPEC+ ministerial response to the UAE departure. Next test of cartel discipline on supply.

• UK Department for Science, Innovation and Technology decision on the Cobalt Park OpenAI site reuse. Any reallocation would signal the AI Growth Zone framework still has industrial momentum despite the OpenAI exit.

I’m building The Brief to be the most useful 5 minutes in your week. If you found it interesting, please forward it. If you did not, send me a message and tell me why.

For enquiries on how we can work together - either for advisory or content (collaborations, podcast, articles), please reach out via dan@danmottram.com.

The stories worth attention.

Rolls-Royce SMR and Great British Energy-Nuclear signed the two-stage contract on Monday, enabling site-specific design and early works to begin immediately at Wylfa on Anglesey. It is the UK’s first small modular reactor programme. The National Wealth Fund has committed up to £599m to Rolls-Royce SMR, on top of £2.6bn already allocated through the 2025 Spending Review. Three 470MW units are targeted to supply the equivalent of 3 million homes from the mid-2030s. Peak construction will support around 3,000 jobs on site and another 5,000 across the wider UK supply chain. Rolls-Royce also holds commitments for up to six further units in the Czech Republic, making it the only SMR developer with multiple European site agreements in place.

The noise around SMR programmes has run well ahead of the orders for two years. Most of the competitive field has slipped on schedule, cost or regulatory approval. The meaningful change this week sits a layer deeper. Long-lead equipment orders are now flowing into a UK nuclear supply chain that has not been active at this scale since Sizewell B in 1995. Pressure vessels, heavy forgings, turbines, control systems and specialist civil works are coming back into the order book of precision-engineering firms that have spent three decades servicing other sectors.

My read: Treat Wylfa as a multi-decade demand signal. The manufacturing pull-through is where the money will be made and the industrial base will be rebuilt. The risk sits entirely in the schedule. Every nuclear programme in British history has been re-paced by politics, and the mid-2030s delivery target gives that vulnerability a long runway. Supply chain firms that pace their investment to regulatory milestones rather than contract signings will end up ahead.

Contract signed for delivery of UK’s first SMRs

ACROSS THE SECTORS

Reuters reported on Thursday that US officials have warned several European governments that contracted weapons deliveries will be delayed because the Iran war is consuming American munitions inventories. Estonia and Lithuania publicly confirmed they had been notified. Foreign Military Sales orders are the most exposed category. The signal is awkward given that the US has simultaneously been pressing European NATO members to increase their share of American equipment in recent procurement rounds.

This is a trust shock more than a logistics problem. European defence ministries planning a decade-long rearmament cannot safely assume that US supply is reliable under wartime load. The second-order effect is a stronger case, already being made inside most European defence establishments, for indigenous munitions, missile-defence and sustainment capacity. That case was harder to argue when American factories were available on time. It has become easier this week.

Forward Prediction: At least three European governments that were intending to expand American procurement in 2026 will restructure or partially onshore those orders by the end of Q3. Watch Poland, Germany and the Nordic bloc for the first moves.

Reuters: Exclusive: US to delay weapons deliveries to some European countries due to Iran war, sources say

Chery told Reuters over the weekend that its next leg of European capacity will come through partnerships with incumbent carmakers rather than greenfield construction. France is on the shortlist as a launch market for the Jaecoo and Omoda brands. The existing Barcelona joint venture with Ebro, on the former Nissan plant, targets 200,000 units per year by 2029 and is already looking undersized against the volumes Chery is shipping. European sales rose almost six-fold last year to 120,147 vehicles. The move echoes EU guidance earlier this year that Chinese EV makers can avoid punitive tariffs by meeting reference prices and demonstrating that state subsidies are not being used to undercut European manufacturers.

The EU is converging on the solar playbook. Chinese capacity gets onto European soil by design, into European supply chains, and into European regulatory scrutiny. European incumbents gain time. Protection is another matter. For Chinese OEMs, the path to volume runs through local plants and underused incumbent capacity, which is faster, cheaper and less politically exposed than greenfield. The risk for European builders with stranded plants is that their own assets become the route-to-market for their competitors.

Reuters: China’s Chery looking to expand car production in Europe

Aurora has tripled its driverless trucking network to 10 routes across the U.S. Sun Belt, accumulated 250,000 driverless miles with zero Aurora Driver-attributed collisions, and has announced observer-free operations on International LT Series trucks in Q2 2026. The company targets more than 200 driverless trucks by year-end and projects positive free cash flow in 2028. The autonomous vehicle sector simultaneously drew regulatory attention this week: Senator Ed Markey sent letters to seven AV companies - including Aurora, Tesla, and Waymo - demanding disclosure of their reliance on remote assistance operators, after a congressional hearing exposed Waymo’s use of remote staff in the Philippines. All seven refused to disclose, prompting Markey to call on NHTSA to investigate and signal legislation to impose guardrails.

Aurora’s operational progress is genuine, but the remote-operator disclosure fight points to a structural vulnerability the sector has not resolved: Level 4 autonomy on public roads still requires human intervention more often than the marketing suggests. Legislation mandating transparency on remote-operator frequency would reprice the labour-cost thesis underpinning autonomous trucking economics and complicate the capital case for fleet customers already making long-term commitments.

StockTitan / Aurora IR: Aurora to Host First Quarter 2026 Business Results

Three separate deployments moved from demonstration to production reference in announcements this week, timed to the opening of Hannover Messe on Monday. At Siemens’ Erlangen electronics factory, UK-based Humanoid and NVIDIA logged 60 tote moves per hour, over eight hours of continuous uptime and a pick-and-place success rate above 90% with the HMND 01 Alpha, built on NVIDIA’s physical AI stack. BMW confirmed a humanoid test deployment at its Leipzig plant, ahead of a full pilot phase starting this summer. In Nanchang, Agibot’s G2 humanoid began working alongside humans on Longcheer Technology’s high-speed tablet production lines, claimed as the first extensive industrial deployment of embodied AI in consumer electronics manufacturing.

The tote-moves-per-hour figure is the obvious metric. More consequential is the seven-month hardware development cycle for the HMND 01, made possible by simulation-first training. That compression is what starts to make the ROI calculable rather than speculative. Hannover Messe opens on Monday with roughly 15 humanoid companies on the exhibition floor and Siemens demonstrating a flexible shoe production line with an autonomous packing robot and a humanoid working alongside it. European manufacturers get four days to decide whether humanoids are a capability they build internally or a vendor-managed bolt-on.

Siemens / PRNewswire: Siemens and Humanoid bring Physical AI to the factory floor

ASML reported Q1 2026 on Wednesday, with net sales of €8.77bn, net income of €2.76bn, and gross margin of 53.0%. Both figures beat consensus. The company raised full-year 2026 revenue guidance to €36–40bn, from €34–39bn. The commentary is what matters. Management confirmed that approximately 19% of system sales in the quarter went to China and that tightening Dutch and US export controls are already baked into the guidance. In Washington, lawmakers scaled back the broadest parts of the proposed MATCH Act under semiconductor-industry lobbying over recent weeks, though they kept a nationwide restriction on ASML’s DUV immersion systems for China and maintained targeted controls on servicing restricted Chinese fabs. A committee vote is expected in the coming days. TSMC’s own Q1 on Thursday, with revenue at $35.9bn and gross margin at 66.2%, confirmed that the capex engine pulling ASML’s order book has lost none of its tension.

AI infrastructure demand is strong enough to absorb policy risk, until it is not. ASML is Europe’s chokepoint asset in the semiconductor stack and an increasingly exposed one. For European industrial users of edge compute and AI accelerators, equipment pricing and availability will remain unpredictable through the next two review cycles. The July 2026 review of US Section 232 datacentre chip classifications is the next real inflection. Any tightening of which chips qualify for tariff exemption will reprice hyperscaler procurement timing and flow through to ASML and TSMC order books inside one quarter.

Investing.com: ASML raises 2026 revenue outlook as AI chip demand drives Q1 beat

THE WEEK AHEAD

• RTX (RTX), Tuesday 21 April, pre-market. After this week’s weapons-delay story, the read on munitions, engine supply-chain strain and European FMS conversion matters more than the EPS print.

• GE Vernova (GEV), Wednesday 22 April, pre-market. GEV stock is up 51% YTD. The number to watch is gas-turbine backlog visibility and Electrification backlog growth tied to European and US data-centre power demand. Any update on the Hitachi nuclear SMR pipeline matters for the European read.

• Texas Instruments (TXN), Wednesday 22 April, post-market. Industrial analog is the cleanest read on factory activity globally. TXN’s channel commentary will be the most reliable signal on whether Q1’s Eurozone PMI expansion is holding into Q2.

• Intel (INTC), Thursday 23 April, post-market. The question is whether server and AI-related demand is strong enough to offset PC weakness, and whether guidance tightens around foundry and advanced-node recovery.

• S&P Global Flash Manufacturing PMIs (UK, Eurozone, US), Wednesday 23 April. Flash Eurozone consensus around 51. A print below 51 would suggest that war-driven cost pressure is starting to bite order books. The UK and US prints are the directional pre-read into the May ISM and the next ECB meeting.

• Hannover Messe 2026, 20–24 April, Hannover. The world’s largest industrial trade fair. 3,000-plus exhibitors, around 15 humanoid companies on the floor, and the new Defence Production Park showing 40 companies on security-critical manufacturing. Chancellor Merz is expected to open the fair with Siemens CEO Roland Busch on Monday. Watch the Defence Production Park. It is the thread most industrial strategists will miss while watching the robots.

• Blue Origin New Glenn-3 launch, today. NG-3 carries AST SpaceMobile’s BlueBird 7 satellite, with booster recovery attempted for the second time. Third mission of the New Glenn programme and a direct test of reusability at cadence.

• US Section 232 semiconductor tariff classification review, due July 2026. Definitions of which datacentre chips qualify for exemption will reprice hyperscaler procurement timing and ASML and TSMC order flow in Q3.

• MATCH Act committee vote, expected in the coming weeks. Any acceleration or further amendment will move ASML and Tokyo Electron directly, and signal where allied export-control policy is settling.

• SpaceX Starship Flight 12, targeting early May after a third delay. Further slippage strengthens the case for commercial launch diversification and puts more pressure on the Blue Origin west-coast pad negotiation.

• Hannover commercial terms, all week. Anyone presenting deployment economics for humanoid robotics or agentic software, rather than on-stage demos, is worth reading in detail. The commercial terms are the point.

ISI. Signal over noise.

A NOTE FROM DAN

This first edition is an extended version covering a few weeks of material I’ve earmarked. Going forward, The Brief will publish every Sunday morning as a weekly roundup of the most strategically meaningful developments across the ISI domains. Signal over noise. Every week.

THE WEEK(S)

Four events. One pattern.

Humans left Earth orbit for the first time in 53 years. The most consequential IPO in a generation was filed. One engine supplier is throttling the world’s largest aircraft manufacturer to 30 fewer deliveries than its production lines can handle. And oil spent the week above $100 before a ceasefire dragged it back to $96.

The pattern is the same across all four: physical constraints are reasserting themselves over financial models. Orbital ambition runs on hardware that takes decades to certify. A $2 trillion valuation is built on a rocket that flies, satellites in orbit, and spectrum rights that cannot be conjured. An aircraft manufacturer with 8.8k aircraft on backorder cannot fill them because one critical-path supplier. And every $10 on the oil price is a hidden industrial tax that quietly improves the relative economics of firm baseload power.

And in case you missed it, the first

ISI Sector Report can be downloaded here ↓

Top 25 OEM Auto Companies

Subscribe now

SIGNAL

The stories worth attention.

Tesla, SpaceX, and xAI unveiled the Terafab semiconductor complex on 22 March from the derelict Seaholm Power Plant in Austin: a planned $20-25bn fabrication facility targeting 2nm process technology and 100,000 wafer starts per month, scaling eventually to 1 million. The scepticism was immediate and well earned. A 1-million-wafer target from a consortium with zero process manufacturing experience was a capital commitment story, not a capability story.

On 7 April, Intel changed that. The formal announcement placed Intel not as a supplier but as the entity that will actually run the process: “Intel is proud to join the Terafab project with SpaceX, xAI, and Tesla to help refactor silicon fab technology. Our ability to design, fabricate, and package ultra-high-performance chips at scale will help accelerate Terafab’s aim to produce 1 TW/year of compute.” INTC surged 11.4% on 8 April on 180 million shares, 64% above average volume, breaking above $60 for the first time in years. The market read the announcement correctly: Terafab is an Intel Foundry expansion with Musk’s companies as anchor customers, not a Tesla greenfield fab. Tesla and SpaceX commit demand. Intel commits the 57 years of manufacturing experience that actually make chips.

My read: Intel has struggled for years to attract meaningful foundry customers. Terafab is the flagship win it needed at exactly the moment it needed one. For TSMC, it is a demand signal. A significant anchor customer group, spanning chips for autonomous vehicles, rockets, and large-scale AI inference, has formally directed volume toward a competitor. The 2027 production target remains aggressive, but the credibility gap no longer exists in the same form.

Bloomberg: Terafab: Intel to partner with Musk companies on chip factory

ACROSS THE SECTORS

SpaceX filed a confidential draft registration statement with the SEC on 1 April, targeting a June 2026 listing at $1.5 to 2 trillion with a $75bn raise. At the lower bound, that is 2.6x the size of Saudi Aramco’s record $29.4bn offering. The company enters the public market as the eighth-largest on earth, behind TSMC and Saudi Aramco, ahead of Tesla at $1.37 trillion.

The filing follows February’s all-stock acquisition of xAI (Elon Musk’s AI company plus X/Twitter) at approximately $250bn. See the previous ISI article from February ↓

Industrial Sector Insights

The full stack: SpaceX x AI

On 2 February 2026, SpaceX announced the acquisition of xAI in an all-stock deal valuing the combined entity at $1.25 trillion. The transaction, the largest in corporate history, folds rockets, satellite broadband, a social network, and a frontier AI lab into a single private company. Days earlier, SpaceX had filed an FCC application to place up to one …

Read more

7 months ago · 5 likes · Dan Mottram

The prospectus will disclose, for the first time, combined revenue and margins across three businesses: Starlink (estimated $15-26bn revenue, $8bn profit), launch services, and AI. Twenty-one underwriting banks have been assembled. Investor briefings are planned for after Easter. Polymarket gives a 94% probability of SpaceX going public before 2027.

This valuation sets comparables for every commercial space company in the sector. It also places Starlink’s economics under public scrutiny: subscriber growth, churn, capital intensity, and the margin profile of a satellite internet business, with no credible competitor at scale.

Bloomberg - SpaceX files confidentially for IPO

The scale of this listing has no precedent in venture-backed exits. SpaceX's projected $1.5 to $2tn valuation would, on its own, exceed the combined exit value of every US VC-backed IPO from 2015 to 2025: $1.22tn total across hundreds of companies. Meta's $104.2bn debut in 2012 was the previous high-water mark. OpenAI and Anthropic, both expected to list within the next 12 to 18 months at $1tn and $500bn respectively, would push the three largest AI-era IPOs past $3tn combined.

That volume of new listings will reshape index composition. SpaceX will almost certainly qualify for S&P 500 inclusion within quarters of listing, subject to meeting the profitability criteria that OpenAI and Anthropic are unlikely to satisfy on debut. All three will rank in the top tier of Nasdaq-100 eligibility by market cap, and Nasdaq's proposed fast-entry rule - triggered for companies ranking in the top 40 constituents at listing - could compress the usual seasoning period to as little as 15 trading days.

Passive funds tracking those indices will be forced buyers at scale, pulling capital from existing holdings to fund the rebalancing. The maths is blunt: domestic equity index funds hold roughly $13.6tn in US equities. Adding companies of this scale as new constituents requires rescaling existing weights downward. For mid-cap industrials and legacy tech names sitting at the margins of index inclusion, the dilution effect is real. Capital does not materialise from nowhere. It reallocates.

NASA launched Artemis II on 1 April, sending four astronauts - Reid Wiseman, Victor Glover, Christina Koch, and CSA’s Jeremy Hansen - on a free-return trajectory around the Moon. On 6 April, the crew broke Apollo 13’s 56-year distance record, reaching a maximum of 252,756 miles from Earth. They passed behind the Moon in a planned 40-minute communications blackout, witnessed a total solar eclipse from lunar orbit, and observed the far side with human eyes for the first time. On 10 April, Orion splashed down off San Diego at 8:07 p.m. EDT after 9 days, 1 hour, and 32 minutes covering 695,081 miles.

Three technical issues flagged for resolution before Artemis III: a helium leak requiring valve system redesign, Avcoat heat shield inspection, and a toilet system fault. Artemis III, now targeting mid-2027, is planned as an Earth-orbit rendezvous and docking test, with Artemis IV targeting early 2028 for the first crewed lunar landing since Apollo 17.

The industrial base is as significant as the mission: Boeing built the SLS core stage, Lockheed Martin built Orion, Airbus built the European Service Module, Northrop Grumman supplied the solid rocket boosters. Artemis II de-risks the crewed lunar landing and opens procurement cycles in cislunar infrastructure, advanced propulsion, and lunar logistics - the contracts that follow successful hardware validation.

Airbus CEO Guillaume Faury confirmed the 2026 delivery target at approximately 870 aircraft despite production capacity exceeding 900 and an order backlog of 8,754 aircraft. The constraint is Pratt & Whitney, which supplies GTF engines for roughly 40% of the A320neo family and has failed to commit to ordered volumes for either 2026 or 2027, agreements that are typically settled 18 months in advance. Faury’s characterisation on the February earnings call was blunt: “They have informed us that they won’t be able to provide the number of engines we were anticipating. We find ourselves in an unsatisfactory position.” In March, Airbus escalated to formal legal action, seeking financial damages over delayed engine shipments.

Q1 2026 delivered 114 aircraft to 46 customers, down 16% year-on-year and the lowest first-quarter figure since 2009, despite March recovering to 60 deliveries after just 19 in January. The 2025 full year came in at 793 aircraft, below the original guidance of 820. Boeing delivered 382 aircraft in 2025, its worst year since 2020. The production ramp target for the A320neo family has been revised down: 70 to 75 aircraft per month by end-2027, stabilising at 75 thereafter, against a prior target of 75 per month by 2026.

The root cause is a 2023 recall of PW1000G engines manufactured with contaminated powdered metal, which has overwhelmed MRO capacity. MRO turnaround times have ballooned from an industry standard of 60 days to nearly 300 days, forcing Pratt to divert new engines intended for Airbus production lines toward maintenance of grounded airline fleets. The engine shortage is not a procurement failure. It is a manufacturing quality problem that Pratt Whitney cannot solve fast enough to serve both its in-service obligations and Airbus’s production ramp simultaneously.

The lesson is structural: Airbus has the orders, the lines, and the materials. It cannot fill aircraft because one critical-path supplier failed. Resilience requires redundancy at the critical-path layer, and with CFM also unable to supply additional engines beyond agreed volumes in 2026, Airbus has no short-term alternative.

Yahoo Finance: Airbus Q1 deliveries fall 16% to 114 jets, lowest since 2009

Tesla delivered 358,023 vehicles against a consensus of 365,645, but the miss is secondary to the inventory signal. Production of 408,386 units against those deliveries created a 50,363-unit gap: the largest single-quarter production-delivery divergence in company history. Energy storage collapsed 38% sequentially to 8.8 GWh. TSLA closed the week at $348.95, down approximately 20% year-to-date from a January high of $448.96.

Two developments since the delivery report shift the narrative context for the 22 April earnings call. First, Cybercab mass production is now under way at Giga Texas, with April marking the official start of the production ramp after initial units rolled off the line in February. Musk’s characterisation is deliberate: “very slow in the beginning,” scaling to very high volume in due course. Second, the Dutch Vehicle Authority granted Tesla the first European type approval for FSD Supervised on 10 April, following 18 months of testing and 1.6 million kilometres driven on EU roads. EU-wide mutual recognition is Tesla’s projection for summer 2026, not a confirmed regulatory outcome.

At Q1’s annualised pace, Tesla would need to average 444,000 deliveries per quarter for the rest of 2026 to meet full-year consensus. The 22 April earnings call must address the inventory build, the energy storage collapse, and the Cybercab ramp simultaneously. The question is whether autonomy rewrites the valuation model before the delivery trend defines it.

Electrek - Tesla Q1 2026 delivery results

Unitree Robotics filed for a $610m IPO on Shanghai’s STAR Market on 20 March, the first profitable humanoid robotics company to go public. Its 2025 prospectus discloses revenue of 1.71bn yuan ($250m), up 335%, with adjusted net profit of 600m yuan ($90m), an 8x increase. The margin story is the signal: gross margins reached nearly 60% despite average selling prices falling from $85,000 in 2023 to $25,000 in 2025, driven by vertical integration across core components. Unitree shipped 5,500 humanoid units in 2025, claiming the top global market share, with plans to reach 75,000 humanoid and 115,000 quadruped units annually within five years. UBTech signed a strategic cooperation agreement with Siemens Digital Industries Software on 16 March, targeting 10,000 industrial humanoid units in 2026, with Siemens providing digital twin and lifecycle manufacturing software across the full production chain.

The honest caveat in the prospectus: industrial pilots remain limited, tour-guide and reception roles account for 50-70% of industry revenue, and current humanoids sustain roughly 90 minutes of continuous operation against the 95-99% uptime real factories require. Unitree is allocating $294m, 48% of IPO proceeds, to embodied AI development, an explicit acknowledgement that the hardware is ahead of the intelligence. China currently leads global humanoid shipments by volume, with the domestic embodied AI market forecast at 400bn yuan ($55bn) by 2030. The first profitable IPO in the sector proves the business model is real. The labour substitution question is still open.

Robotics and Automation News: Unitree Robotics files for $610m IPO as humanoid robot sales surge

Two deals defined the emerging industrial robotics consolidation. SoftBank agreed to acquire ABB’s entire robotics division for $5.375bn, a deal expected to close mid-to-late 2026 pending regulatory approvals in the US, Europe, and China. ABB’s robotics business generated $2.3bn in 2024 revenues at a 12% margin. SoftBank’s stated rationale is Physical AI: fusing AI with real-world robotics at industrial scale. Separately, HII, America’s largest shipbuilder with 44,000 employees, signed an MOU with GrayMatter Robotics to integrate Physical AI into hull surface preparation, coating, and inspection. HII’s throughput increased 14% in 2025 under its High-Yield Production Robotics initiative and targets a further 15% in 2026. The US Navy’s shipbuilding backlog is the most acute military-industrial capacity constraint in the country. Two consecutive years of double-digit throughput gains from autonomous systems confirm this is a production lever, not a technology showcase.

GlobeNewswire: HII teams with GrayMatter Robotics to integrate Physical AI

The Eurozone Manufacturing PMI rose to 51.6 in March 2026 from 50.8 in February, its strongest reading in 45 months, with new orders expanding and six of eight surveyed economies in growth territory. Germany returned to manufacturing growth for the first time in three and a half years, with input costs rising at their fastest pace in 38 months as energy price pressures fed through. The timing is awkward. The Turnberry trade deal (the framework agreed in February 2026 between the EU and the Trump administration) locks in a 15% tariff rate for EU exports to the US, but as competing nations such as China (31.1% effective tariff rate in 2025) and India (20.5%) see their tariff rates converge toward 15% in 2026, Europe’s relative competitive position on the US market deteriorates. ING Research estimates US tariffs will reduce EU-US exports by approximately 4.6% in 2026. The manufacturing recovery is real. The tariff headwind into its strongest market arrives at the worst possible moment.

Trading Econ/S&P Global: Euro area factory activity surprises on the upside

Samsung Electronics’ unions, representing approximately 90,000 workers, voted 93.1% in favour of an 18-day general strike from 21 May to 7 June, following the collapse of three months of wage negotiations. The central demand is the removal of Samsung’s performance bonus cap, a concession rival SK Hynix already made - a distinction that is now driving employee attrition from Samsung toward both SK Hynix and, according to the union leader, Tesla. The timing is the risk. The strike window overlaps precisely with AMD and Meta auditing Samsung’s HBM4 production line as part of the $100bn MI450 GPU programme, in a market where SK Hynix already controls approximately 70% of HBM allocation. Samsung management’s own statement is telling: even a single strike “could undermine customer confidence and take years to restore.” A rally at Pyeongtaek on 23 April is the next escalation point. No agreement has been reached since the vote.

The broader context is the worst possible backdrop for a supply disruption. Gartner now projects global semiconductor revenue will exceed $1.3 trillion in 2026, up 64% year-on-year and the highest growth rate in two decades, with memory revenue alone forecast to triple. DRAM annual prices are forecast up 125% in 2026, with no meaningful relief expected until late 2027. Texas Instruments raised some product lines by up to 85% on 1 April and Infineon raised power switch prices 5 to 15%. The supply chain is tightening from multiple directions simultaneously. A Samsung strike would not cause the shortage - it would deepen one already in progress.

Xinhua / Reuters: Samsung workers approve strike plan, raising chip supply concerns

Despite $650bn+ committed by hyperscalers to AI infrastructure in 2026, nearly half of planned US data centre projects face delays or cancellation due to electrical equipment shortages. Custom transformer lead times have extended to 52-86 weeks at best, and in some cases five years. Two companies demonstrated the circumvention logic: Crusoe is manufacturing its own switchgear and building a dedicated 900 MW on-site power plant for its 2.1 GW Abilene campus serving Microsoft. Meta has committed to building ten dedicated gas-fired plants totalling 7,400 MW, more than 60% of Entergy Louisiana’s entire current generation capacity, as the power bridge for its $27bn Hyperion data centre complex, with nuclear PPAs from Oklo, TerraPower, and Vistra bridging to 2030-2035. The AI buildout has hit the atoms wall. Companies that own their power generation and manufacturing infrastructure are structurally advantaged.

Crusoe: Crusoe announces new 900 MW AI factory campus in Abilene, Texas

The Strait of Hormuz crisis compressed six weeks of geopolitical risk into a single price chart. Brent reached approximately $113-114/bbl before the 7 April US-Iran ceasefire sent it crashing to $93.76 on 8 April. The Islamabad negotiations that followed lasted 21 hours before collapsing on 12 April, closing the week near $96.69/bbl. The ceasefire window expires approximately 22 April. Two European nuclear milestones confirmed the structural direction. Rolls-Royce SMR was selected for both Wylfa, UK (up to 1.5 GW, with development commencing in 2026 and first power targeted for the mid-2030s) and Czech utility CEZ (up to 3 GW). The Wylfa selection positions the UK as the first European country to implement modular nuclear power plants at grid scale, with the UK government investing £2.5bn as the foundation for a fleet development strategy targeting 24 GW of nuclear capacity by 2050. In Canada, Ontario Power Generation applied for a 20-year operating licence for the first BWRX-300 SMR in a G7 country at Darlington - a major regulatory milestone for the GE Hitachi design that will inform timelines across Europe and the US. The structural argument for firm power does not require a crisis. It requires only the data centre capex charts.

UK sets stage for nuclear rebuild with ambitious 2026 expansion plans

THE WEEK AHEAD

• ASML (ASML), Wednesday 15 April. Q1 2026 results, guidance €8.2-8.9bn, EPS consensus $7.68. The read on EUV order book health and customer capacity planning signals whether the semiconductor super-cycle is holding.

• TSMC (TSM), Thursday 16 April. Full Q1 call after the revenue beat at $35.7bn. Q2 guidance and capex commentary on the $52-56bn 2026 plan are the numbers that matter.

• Airbus pre-quarterly communication, Tuesday 15 April. Ahead of the full Q1 2026 results on 28 April. Delivery rate and Airbus Defence and Space order momentum are the primary data points for European aerospace supply chains.

• Lockheed Martin (LMT), Monday 21 April. First major defence prime earnings. F-35 programme delivery cadence and any Hormuz supply-chain commentary from its aerospace manufacturing operations.

• Tesla (TSLA), Tuesday 22 April. Consensus EPS of $0.24 on revenue of ~$22.8bn. The 50,363-unit production-delivery gap, energy storage collapse, FSD 14.3 European approval, and Cybercab ramp update must all be addressed. Any Terafab or Intel partnership commentary moves the stock.

• Intel (INTC), Wednesday 23 April. First results since the Terafab announcement and the Fab 34 Ireland repurchase from Apollo for $14.2bn. Foundry margin and customer pipeline commentary will determine whether the market’s 11.4% reaction was justified.

• RTX (RTX), late April. Full-year sales guidance of $92-93bn. Pratt & Whitney engine delivery cadence and any tariff impact quantification.

• Ceasefire expiry, approximately 22 April. Islamabad talks collapsed on 12 April. If negotiations do not resume, Brent reprices toward $110+ on breakdown risk. Every industrial sector with energy input costs watches this date.

• IMF-World Bank Spring Meetings, from Monday 13 April. Global trade and growth forecasts in a tariff-disrupted environment will set the context for industrial capex decisions through H2 2026.

• UK Spending Review, June 2026. The June review sets defence departmental budgets through 2028-29. The gap between Starmer’s 3.5% GDP commitment and the OBR’s current forecast path is the clearest near-term funding constraint for UK defence primes and their supply chains.

• Blue Origin New Glenn launch, Thursday 16 April. Orbital payload test. The commercial launch market share question is live.

• Samsung union rally at Pyeongtaek, 23 April. No agreement reached since the 93.1% strike vote. An 18-day general strike starting 21 May would disrupt HBM4 production at the precise moment AMD-Meta volumes are audited.

• SpaceX S-1 prospectus, expected late April or early May. The public filing reveals the combined financials of SpaceX, Starlink, and xAI for the first time. Every assumption about the Musk empire gets tested in one document.

• Rolls-Royce SMR programme milestones, ongoing. Site licensing and environmental permitting at Wylfa is under way. Amentum selected as programme delivery partner in January 2026. The first European grid-scale SMR programme is in execution, not planning.

The global automotive industry generated over $2.5 trillion in combined revenue in 2025. It also wrote off $60 billion in electrification investments in the same year.

GM abandoned its all-electric-by-2035 target and took a $7.6 billion charge. The stock rose 4% on the announcement day, and surged a further 9% on earnings day as investors supported the pivot.

Ford’s Model E division lost $4.8 billion, Volkswagen’s operating income fell 53.5%, and Stellantis posted a €22.3B loss, driven by €25.4B in exceptional charges.

Meanwhile, BYD overtook Tesla as the global BEV sales leader with 2.26 million units, selling at an average price roughly half of Tesla’s.

Toyota quietly validated the hybrid strategy everyone dismissed, delivering 4.4 million HEVs at healthy margins while the rest of the industry burned capital chasing battery electric volume.

Ferrari’s EBITDA margin hit 38.8%. The next closest full-line manufacturer operates in a different economic universe.

So what’s actually happening inside this industry?

I broke down all 25 of the world’s most valuable automotive companies, examining their financial positions, strategic pivots, and the tensions their boards cannot defer.

Five patterns emerged: the market rewards story over substance, hybrid was the answer nobody planned for, China won the technology race (and traded it for margin), Europe is retreating into luxury, and India is the next growth engine.

This deep dive maps all of it, and you can read it in the attached PDF at the end of this post.

In future months, I will publish deep dives into other sectors of the industrial economy. These will cover nine domains, and where they converge.

1. Aerospace, Defence & Space

2. Mobility & Transportation

3. Heavy Machinery & Capital Goods

4. Robotics & Embodied AI

5. Advanced Manufacturing & Materials

6. Software & the Agentic Factory

7. Semiconductors & Electronics

8. Data Centres & Digital Infrastructure

9. Energy, Nuclear & Grid

If enough people are interested and subscribe, I’ll take these further and strive to make them a valuable, insightful asset for you. I will also look to you to influence the future research topics and share insights with others.

I hope you enjoy reading this, and do let me know what you think in the comments below or via message.

Sector Report PDF below ↓

Automotive OEM Sector Report V1 26 03 25

29.4MB ∙ PDF file

Download

Download

Schneider Electric’s FY2025 results are the kind that invite lazy superlatives.

Revenue crossed €40 billion for the first time. Free cash flow hit €4.6 billion. Return on capital employed (ROCE) cleared 15%. The backlog stands at €25.4 billion, up 18% year-on-year.

The headline numbers are strong.

The structural story underneath them is stronger.

Schneider is building a compounding loop. Physical infrastructure generates operational data. That data feeds a growing software and services layer. The software makes the installed base stickier and more valuable. Management calls it the Digital Flywheel. In FY2025, it accounted for 62% of group revenue, up from 57% a year ago, growing at 15% organically. The target is 70% by 2030.

The company that started as a French industrial conglomerate in steel and mining is now the closest thing the electrical infrastructure market has to a platform business.

The machine today

Schneider reports through two segments: Energy Management and Industrial Automation. The split is roughly 80/20, but the real business model is better understood through its product architecture: Products (47% of revenue), Systems (34%), and Software & Services (19%).

Energy Management is the engine. It grew 10.3% organically in FY2025, with margin at 21.8%. The backlog here reached €21.3 billion, up 21%, driven overwhelmingly by data centre and grid infrastructure demand. Q4 alone saw 11% organic growth, with data centre orders accelerating to triple-digit growth year-on-year in the pure hyperscale and colocation segment.

Industrial Automation is recovering. After a prolonged destocking cycle in discrete automation, the segment returned to positive growth: 3% for the full year, accelerating to 8% in Q4. AVEVA, the software platform acquired in 2023, is the crown jewel here, delivering double-digit growth with recurring revenue at 85% and ARR growth of 12%.

The geographic story reinforces the structural thesis. North America is the largest region at 39% of revenue, growing 15% for the year and nearly 18% in Q4. Data centres are the primary driver, but grid infrastructure spending and reshoring investment are contributing meaningfully. Schneider has 83% local manufacturing content in North America, a deliberate decision over the past decade that now provides near-complete insulation from tariff exposure.

The margin paradox

The most instructive data point in the FY2025 results is one that looks like bad news: gross margin fell 40 basis points organically.

The explanation reveals the strategy. Systems, the fastest-growing revenue line at 19% Q4 growth, carry lower gross margins than standalone Products. Raw material costs (copper, steel, silver, rare earths) added further pressure. And the mix is shifting deliberately toward higher-volume, lower-margin system sales that drive recurring software and services revenue downstream.

The resolution sits in support function costs (SFC). Schneider cut its SFC ratio from 24.1% to 23.3%, a 90 basis point organic improvement driven by scale leverage and operational efficiency. The combination of negative gross margin and strong SFC leverage delivered a net 50 basis points of adjusted EBITA margin expansion, reaching 18.7%.

This is a business choosing to grow the top of the flywheel (systems) while extracting operating leverage from the middle (SFC). The gross margin compression is the cost of building the installed base that funds the software moat.

The data centre gravity well

Data centres are Schneider’s highest-growth, highest-visibility end-market. The pure data centre segment delivered triple-digit order growth in Q4, and the backlog provides 18 to 24 months of revenue visibility.

Three dynamics make this positioning distinctive.

First, scope. Schneider supplies the full electrical and thermal stack: medium-voltage switchgear, UPS systems, busway distribution, rack-level PDUs, and cooling. The Motivair acquisition added liquid cooling. EcoStruxure software manages it all. The NVIDIA “grid-to-chip” partnership is emblematic: Schneider is designing the reference architecture for next-generation AI data centres, including 800V DC distribution.

Second, the hyperscaler capex wave. Combined capital expenditure from the five largest cloud and AI infrastructure providers is projected at $630 billion for 2026. Every dollar of that spend requires electrical and thermal infrastructure. Schneider’s backlog growth, concentrated heavily in North America, reflects contracted demand against this wave.

Third, stickiness. Once EcoStruxure manages a data centre’s power and cooling infrastructure, switching costs are substantial. The software monitors, optimises, and predicts across the installed hardware base. Each new rack, each new building, each new campus deepens the integration. The hardware sale is the acquisition channel for a long-duration software relationship.

Building the regional moat

Schneider’s multi-hub manufacturing strategy is accelerating. The principle is simple: manufacture 90% of what you sell within the region where you sell it. In North America, the figure is already 83%.

The India play is the boldest recent move. In December 2025, Schneider completed the acquisition of the remaining 35% of Schneider Electric India Private Limited (SEIPL) from Temasek for approximately €5.5 billion, financed primarily through a €3.5 billion EMTN bond and a separate €750 million convertible bond (OCEANE). India is now a fully owned strategic hub: a manufacturing base for the international region, a fast-growing domestic market, and an innovation centre. A new liquid cooling factory and the Bengaluru campus expansion reinforce the commitment.

In the US, $700 million of additional investment through 2027 targets data centre capacity expansion: prefabricated power modules, liquid cooling systems, and EcoStruxure software development. New facilities in Mt Juliet (Tennessee) and Macon (Europe) are already operational.

The regionalisation strategy serves three purposes simultaneously: tariff resilience, speed to customer, and local innovation. The companies that built long, thin supply chains over the past two decades are now rebuilding them as regional networks. Schneider started earlier than most.

The FX drag and what it reveals

Foreign exchange delivered a €482 million headwind to adjusted EBITA in FY2025, roughly 50 basis points of margin. The primary culprits were a weakening US dollar, Indian rupee, and Chinese yuan against the euro.

For FY2026, Schneider guides a revenue FX headwind of €850 million to €950 million, with a further 10 basis points of EBITA margin pressure.

The FX exposure is significant. It is also a direct consequence of the geographic diversification strategy. A company with 39% of revenue in North America, 26% in Asia-Pacific, and meaningful exposure to India, China, and emerging markets will face translation effects when the euro strengthens. The alternative, concentrating in the eurozone, would eliminate the FX drag and forfeit the growth.

Guidance and what it implies

Schneider targets FY2026 adjusted EBITA growth of 10% to 15% organically, built on 7% to 10% revenue growth and 50 to 80 basis points of margin expansion. The implied margin range is 19.1% to 19.4%.

Gross margin is expected to turn positive during 2026 as pricing actions catch up with raw material and tariff costs. All four regions are expected to contribute to growth, with the US and India leading.

The medium-term targets from the Capital Markets Day reinforce the trajectory. Revenue CAGR of 7% to 10% through 2030. Cumulative organic EBITA margin expansion of 250 basis points. Cash conversion at approximately 100%. If achieved, Schneider would exit 2030 with an EBITA margin above 21% on revenue approaching €55 billion to €60 billion.

The transition beneath the results

A CFO transition is underway. Hilary Maxson departs on 5 April 2026, replaced by Nathan Fast, currently Head of Investor Relations. Management has emphasised that all strategic assumptions and targets from the Capital Markets Day remain unchanged. The succession appears planned and orderly.

The sustainability programme concluded on target: 862 million tonnes of CO2 saved or avoided since 2018, against an 800 million tonne target. A new 2030 roadmap targets 1.5 gigatonnes. Schneider was ranked the world’s most sustainable company by Statista for the second consecutive year. The sustainability credentials are a strong leg to the value proposition for enterprise customers with decarbonisation commitments and a differentiator in public procurement.

Schneider Electric’s FY2025 is the cleanest articulation yet of the infrastructure intelligence thesis. The physical installed base is the acquisition channel. The software layer is the retention mechanism. The data generated in between is the competitive moat that deepens with every system sold.

The gross margin compression is the signal, not the noise.

Schneider is investing in system density to build an installed base that generates recurring, high-margin software revenue for decades. The €25 billion backlog provides visibility. The 111% cash conversion provides optionality. The Digital Flywheel, at 62% of revenue and growing, provides the compounding mechanism.

The question for FY2026 and beyond is execution pace: can the software layer grow fast enough to convert the physical lead into a durable platform advantage before competitors close the gap? The backlog and the AVEVA ARR trajectory suggest the answer is forming.