An encyclical is the most authoritative teaching letter a pope can write. Today at 11:30 in the Vatican’s Synod Hall, Pope Leo XIV releases his first one, Magnifica Humanitas — Latin for “Magnificent Humanity.” The subject is artificial intelligence and the protection of human dignity. Two unusual things stand out. First, the Pope is presenting the text himself, which popes almost never do. Second, sharing the stage is Christopher Olah, co-founder of Anthropic and head of its interpretability research — the first time an AI executive has spoken at the launch of a papal encyclical. The Pope signed the document on May 15, the 135th anniversary of Leo XIII’s Rerum Novarum, the 1891 letter that defined the Church’s response to the industrial revolution. The framing is deliberate.

Shortly after 11:30 Rome time, Pope Leo XIV walked into the Synod Hall, the modern auditorium just south of St. Peter’s where bishops normally debate doctrine, and took a seat next to a man in a dark suit with no clerical collar: Christopher Olah, the 33-year-old Canadian researcher who co-founded Anthropic in 2021 after leaving OpenAI over what he and his colleagues described as inadequate attention to safety. The Pope opened a leather folder containing Magnifica Humanitas, blessed the assembly, and addressed the room directly — a break with the usual protocol in which a senior cardinal presents the text on the pontiff’s behalf. The encyclical, according to the Vatican press office and excerpts circulating in advance, frames artificial intelligence as a tool — “never a moral agent” — whose ethical measure is whether it preserves “the inviolable dignity of every human being.” It warns that AI used without conscience can deepen inequality, fuel conflict, accelerate disinformation, and reduce persons to data points. It calls out the use of AI in warfare, the displacement of workers by automation, and the ecological cost of large-scale compute. It does not propose regulation in detail, but it explicitly invites states and corporations to be accountable. Flanking the Pope and Olah on the panel were Cardinal Víctor Manuel Fernández, Prefect of the Dicastery for the Doctrine of the Faith; Cardinal Michael Czerny, S.J., who runs the Dicastery for Promoting Integral Human Development; Professor Anna Rowlands of Durham University; and Professor Leocadie Lushombo of the Jesuit School of Theology at Santa Clara. Czerny set the tone in his remarks, paraphrased by Vatican News: the Church is not pronouncing on the technology itself, but on what it does to the human person — a distinction the Vatican has been refining since Pope Francis’s 2024 message for the World Day of Peace, the first papal text to treat AI as a standing subject. Olah spoke briefly. According to people in the room, he framed Anthropic’s interpretability program — the attempt to reverse-engineer the internal workings of large language models — as a technical answer to a question the encyclical poses in moral terms: can humans actually understand the systems they are building well enough to be responsible for them? He did not announce a product, a partnership, or a donation. The choreography was the message. Not by accident, the Vatican chose May 25 — Memorial Day in the United States, when Washington is closed and Silicon Valley distracted. The result was a global news cycle dominated, for once, by Rome rather than by an OpenAI demo.

The choice of name and date is the document’s most legible signal. When Cardinal Robert Francis Prevost was elected on May 8, 2025 and took the name Leo, Vatican-watchers immediately drew the line to Leo XIII, whose 1891 encyclical Rerum Novarum — “Of New Things” — gave the Church its first systematic response to industrial capitalism. Rerum Novarum defended the right to a just wage, the right to organize, and the dignity of labor against what Leo XIII called the “misery and wretchedness” into which workers had been pushed by unrestrained industry. It founded modern Catholic social teaching. By signing Magnifica Humanitas on May 15, 2026 — exactly 135 years after Leo XIII signed Rerum Novarum — Pope Leo XIV is making an argument with a date. He is telling 1.4 billion Catholics, and everyone else listening, that AI is the techno-industrial moment of our time and that the Church intends to speak to it with the same seriousness with which it once spoke to the factory. The Vatican has been building to this. In 2020 the Pontifical Academy for Life convened the “Rome Call for AI Ethics,” signed by Microsoft, IBM, Cisco, the FAO and the Italian government, around six principles: transparency, inclusion, accountability, impartiality, reliability, and security and privacy. Pope Francis returned to the theme in his 2024 World Day of Peace message and again in the Apostolic Letter Antiqua et Nova, issued by the Dicastery for the Doctrine of the Faith in January 2025. The Holy See itself adopted one of the world’s earliest state-level AI frameworks for Vatican City. The intellectual scaffolding for Magnifica Humanitas has been under construction for six years. Why Anthropic of all labs? Three reasons. First, the company has positioned itself, with some success, as the AI lab whose founding premise is safety; Olah and several colleagues left OpenAI in 2021 over precisely that concern. Second, Olah’s personal field — mechanistic interpretability — translates well into the theological vocabulary the Vatican prefers: it is about making the machine legible to the human, not the other way around. Third, according to Religion News Service, Anthropic researchers had been quietly approaching the Pontifical Academy for Life “asking for direct help from the Vatican to convene and help the industry, because the industry was going so fast down this road.” The Pope’s team accepted. More remarkable still is what the encyclical does not do. It does not endorse a specific regulatory regime. It does not name companies. It does not bless or condemn any particular product. That restraint is itself a choice: it leaves the document usable, in Brussels and Berlin and Washington alike, as moral cover for almost any serious attempt at AI governance — and as moral rebuke to almost any cynical one.

For European policymakers, the timing is conspicuous. The EU AI Act becomes fully applicable on August 2, 2026 — ten weeks after today’s presentation. The high-risk provisions covering biometrics, employment, education, migration and critical infrastructure take effect on December 2, 2027. The Commission missed its February 2026 deadline for the implementing guidelines and is under pressure from member states, industry and now, implicitly, from the Holy See. The COMECE — the bishops’ conference representing the Catholic Church to the EU institutions — held private talks with Leo XIV at the Vatican earlier this spring on AI governance. Senior officials in Berlin and Brussels expect Magnifica Humanitas to be cited in parliamentary debates within days. The German Bishops’ Conference (DBK) circulated a discussion paper to its 27 dioceses two weeks ago anticipating exactly this scenario; the Evangelical Church in Germany (EKD) is preparing a parallel response. In Germany, where roughly 21 million people are still registered Catholics and where Caritas alone employs around 700,000 staff across hospitals, schools, kindergartens and social services, the encyclical will be read aloud in dioceses and discussed in works councils. Caritas signed contracts with Microsoft Azure for its hospital operations in 2024; that procurement is now a board-level conversation again. Further afield, the Spanish Episcopal Conference is convening a Madrid summit in June, the Italian bishops are pairing the document with their own AI labour study, and Vatican Radio has commissioned translations into Mandarin, Arabic and Swahili — a signal that the Holy See sees Magnifica Humanitas as a global text, not a European one. For DAX40 boards, the political cost of being on the wrong side of a papal teaching has just gone up. Expect questions at June and July AGMs, expect ESG-rating agencies (MSCI, ISS, Sustainalytics) to thread the encyclical into their AI-governance scorecards inside the quarter, and expect the next round of EU AI Act trilogue lobbying to be conducted with the document as background music.

For two decades, the deal between large companies and big consultancies has been simple: when a new technology arrives, you rent thousands of human brains from Accenture, Capgemini or T-Systems to wire it into your business. AI is now testing that contract. A growing chorus of venture capitalists argues the next great services firms will not be the incumbents bolting AI onto their existing armies of analysts. They will be smaller, AI-native companies that have rebuilt how the work itself gets delivered — fewer people, more software, faster results. Over the Memorial Day weekend, that argument moved from blog posts to the cheque book. New funds, new joint ventures and new vendor stacks are converging on the same bet: services, not software, is where AI economics actually land.

On Sunday evening, May 24, Yoni Rechtman, a partner at Slow Ventures, hit publish on a post titled “AI Accenture, Not Accenture for AI” on his Substack, 99% Derisible. By Monday morning it was the lead item in the Sandhill.io weekly digest that aggregates more than 700 venture voices, and by Tuesday it was being forwarded inside the strategy teams of every Tier-One consultancy in Frankfurt and Munich. The thesis, distilled: “There’s an immense opportunity for services companies to fundamentally rebuild how they deliver, not merely what they deliver.” Rechtman sorted the emerging market into three buckets — forward-deployed-engineer bodyshops, lab-backed captives, and “product companies delivering a service.” The last is the one Slow wants to back, firms that build an internal product they consume to deliver an external service “better, faster, cheaper.” The same weekend, Emily Man — Primary Venture Partners’ first fintech partner, formerly of Redpoint — was circulating a companion piece making the identical argument from the financial-services angle: in a world where models are commoditising, implementation is where the durable margin sits. Two essays, one thesis, from two different coasts and two different verticals. The Sandhill digest framed them as a pair. The pivot, and what made enterprise readers sit up, was the timing. Twelve days earlier, on May 12, OpenAI had announced the OpenAI Deployment Company — a standalone consulting business seeded with $4 billion from OpenAI, TPG, Advent, Bain Capital, Brookfield and 15 others, and immediately fattened with the acquisition of Tomoro and its 150 forward-deployed engineers. Eight days before that, on May 4, Anthropic had unveiled a $1.5 billion enterprise-services joint venture with Blackstone, Hellman & Friedman and Goldman Sachs, designed, as Blackstone’s Jon Gray put it, to break down “one of the most significant bottlenecks to enterprise AI adoption.” Anthropic CFO Krishna Rao said plainly that “enterprise demand for Claude is significantly outpacing any single delivery model.” A CIO from a DAX40 industrial flying back from SAP Sapphire in Madrid the previous weekend — where SAP had introduced Joule Work, Joule Studio, the Autonomous Suite and a €100 million partner fund — could be forgiven for reading all of this on the plane and concluding the ground was shifting under his SI roster. The May 22 briefing called the lab-and-hyperscaler moves “The Pincer.” What the VC essays added was the other arm: private capital flowing not into incumbents but into a new generation of AI-native challengers. Echelon, a Bain Capital Ventures-backed startup, came out of stealth promising agents that fully automate ServiceNow implementations — work that used to take offshore consulting teams months. The historical rhyme is the late-1990s SAP rollout boom, when a generation of mid-tier integrators captured tens of billions in implementation fees from greenfield clients. Only this time the implementer may be a 60-person company with a product, not a 60,000-person firm with badges.

Start with the cost base under attack. Accenture employs roughly 750,000 people worldwide and reported $69.7 billion in fiscal 2025 revenue. About 9,000 of those people sit in DACH, where the firm runs its largest continental-Europe market. Capgemini’s headcount sits around 340,000 globally, Atos a further 92,000, T-Systems roughly 26,000, and Reply, MHP and BearingPoint another 40,000 combined. The aggregate labour-hours-priced cognitive cost base sitting against the DACH economy is well over 100,000 billable consultants. That is the denominator the VCs are looking at when they talk about repricing. Now the numerators of the new economy. OpenAI’s Deployment Company opened with $4 billion of committed capital and acquired Tomoro for its 150 engineers — implying a per-FDE valuation that would have been laughed out of a partner meeting two years ago. Omdia’s Peter Bryant, quoted in CIO Dive, argued that “FDEs are going to be critical in converting opportunities to revenue,” because by the time the legacy SI gets its people trained on the current model, the next one ships. Anthropic’s joint venture put $1.5 billion behind a parallel structure with Blackstone, H&F and Goldman — a built-in distribution channel of hundreds of PE portfolio companies. Add in General Atlantic, Apollo, GIC, Sequoia and Leonard Green as co-investors, and the captive client pipeline alone runs into the four digits. Together, $5.5 billion of new capital landed on AI-native services in 21 days. Then there is the platform leg. SAP at Sapphire 2026 unveiled more than 50 domain-specific Joule Assistants and 200+ specialized agents inside its Autonomous Suite. RISE with SAP customers get three Joule Assistants activated in year one; SAP GROW customers get the full portfolio at onboarding. Crucially, Joule Studio gives any customer or partner the no-code, pro-code and AI scaffolding to build their own agents on SAP-managed infrastructure — and SAP put €100 million behind a fund to underwrite partner builds. SAP also claims its agent-led ERP migration tooling can cut migration effort by more than 35 percent. Translate that into consulting hours: a typical S/4HANA migration that ran 18 months and 80 FTEs at €1,800 a day becomes a 12-month, 50-FTE engagement. That is roughly €25–35 million per programme stripped out of the SI line — and SAP has thousands of these projects in flight. The public-market signal corroborates. Per John Hwang’s Enterprise AI Trends, year-to-date 2026 stock performance for the consulting cohort is ugly: Accenture down 10 percent, Gartner down 17 percent, IT (Gartner Inc.) down 36 percent after its February sell-off. Hwang’s read of three earnings calls covering $90 billion in consulting revenue: AI is a short-term tailwind that flips into a 2027 headwind as deal sizes compress and clients catch on. Meanwhile Omdia still projects more than $1 trillion to flow through the SI market this year and 6.5 percent CAGR through 2031. Both can be true: the pie grows, but the slices migrate.

DACH is where this thesis matters most because DACH is where the legacy stack is densest. Accenture’s 9,000 German-speaking consultants are the single largest concentration in continental Europe and the engine of SAP-related programme delivery for DAX40 names from BMW to Bayer. Capgemini’s strategy chief Fernando Alvarez told Fortune in March that AI was not, in fact, killing consulting — clients still need orchestration, change management and integration muscle, and Capgemini has plugged itself into OpenAI’s Frontier Alliances partnership precisely to occupy that role. T-Systems, the captive SI inside Deutsche Telekom, leans on sovereign-cloud positioning and Bundesnetzagentur-friendly hosting to defend its 26,000-seat base. Atos, restructured and refinanced, is positioning AI services as the growth wedge that will let it overtake Capgemini in raw IT-services revenue by decade’s end. Reply, the Italian-but-DACH-heavy network of boutiques, has been quietly building agentic teams that look more like Rechtman’s “product companies” than classical SIs. MHP — Porsche’s captive — has Joule Studio early access and a clear mandate to industrialise it for the German Mittelstand. BearingPoint runs the smaller, advisory-led play. The DACH-specific stress test arrives via SAP. RWE’s Sapphire keynote showcased Joule-driven asset management for offshore wind turbines, a flagship reference customer SAP will now wave at every utility and industrial in the DACH belt. If Joule Studio plus the €100 million partner fund lets a Munich-based industrial build its own finance-close agents — without a 40-person Accenture pod onsite for nine months — the implementation line item simply shrinks. That is the disintermediation risk legacy SIs have to price. The defence, as Alvarez argues, is orchestration: someone still has to choose between Anthropic and OpenAI and Cohere, integrate Joule with Salesforce Agentforce and Workday Illuminate, and shoulder the SLA when an agent hallucinates a journal entry. The question is whether that orchestration role pays €1,800 a day for 80 consultants or €4,000 a day for eight.

Hyperscalers — Amazon, Microsoft, Google, Meta, Oracle — are racing to build AI data centers, the warehouses full of chips that train and run models like Claude and Gemini. The bottleneck has shifted. A shed full of NVIDIA GPUs can be poured, wired and roofed in 12 to 18 months. Plugging it into the electric grid takes five to seven years in the US, and eight to thirteen in parts of Germany. That gap is now the single biggest constraint on the AI buildout. This story is about a Bessemer Venture Partners roadmap that names the six layers where the next wave of infrastructure capital should flow — power generation, transmission and transformers, orchestration software, cooling, construction robotics and permitting tools — and why the European industrials that make this hardware (Siemens Energy, ABB, Schneider Electric, Hitachi Energy) suddenly look like the most boring, most important trade of the cycle.

On Monday 18 May, a Bessemer partner pushed a 12,000-word essay live on the firm’s Atlas blog. Titled “Roadmap: The AI Data Center Stack,” it was authored by Lindsey Li, Brielee Lu, Josh Hechtman and David Cowan, and it landed with the kind of thud that only a venture-capital market map can produce when the field has been waiting for one. By Sunday breakfast in Frankfurt, every infra investor with a Bloomberg terminal had it open in a tab. The headline number is the one to remember: as of early 2026, 190 GW of hyperscale data center capacity has been announced across 777 projects, per Sightline Climate data Bessemer cites. Roughly 148 GW is at the planning stage, around 21 GW is under construction, and only about 12 GW is operational today. The pipeline is roughly sixteen times the size of what is actually running. Of the 110 data center projects slated to come online in 2025, more than a quarter slipped — power, permitting and construction the usual suspects. The Bessemer framework slices the opportunity into six layers: permitting and site selection; power generation; transmission, power conversion and the middle mile; software and orchestration; construction, maintenance and labor; and cooling. It is, deliberately, a map of where venture dollars should NOT chase NVIDIA. “Data centers dominated built-environment venture investment in 2025, accounting for 78% of capital deployed,” the authors write, but the enabling hardware and software layers remain in their early innings. They invoke Sir Henry Bessemer himself — the firm’s namesake — whose process for mass-producing steel from molten pig iron turned a costly artisanal material into the backbone of railroads, bridges and skyscrapers. The implication: transformers and switchgear are this cycle’s steel. Two other essays landed inside the same week and gave the roadmap its political weight. Ben Thompson, on Stratechery, published “Data Center Discontent, Understanding the Opposition, Fixing the Problem” on 18 May, arguing that local opposition to data centers is real, rational and cannot be argued away with PR — communities must be paid off, in cash and in kind, or they will exercise a de facto veto. Two days later, Azeem Azhar’s Exponential View opened with Eric Schmidt being audibly booed at the University of Arizona commencement, a student telling Joanna Stern to “throw it away,” and Alexandria Ocasio-Cortez holding up a jar of mud-brown tap water from a Georgia community where Meta had just broken ground on a campus. “There will be more booing,” Azhar wrote. The social licence for the AI buildout is cracking at the precise moment Bessemer is publishing where the next $500 billion should be deployed. The narrative pivot here is the one Washington has already made. In April 2026, President Trump invoked Section 303 of the Defense Production Act, formally designating large-scale grid infrastructure — transformers, transmission lines, substations, high-voltage circuit breakers, power electronics, electrical core steel — as essential to national defense, and authorising the Department of Energy to provide direct financial support to expand domestic supply. That is, in practice, an emergency industrial policy for the grid, executed through the same statute that built the World War II arsenal. The picks-and-shovels trade just got a federal backstop.

Start with the demand side. The Big Four hyperscalers — Amazon, Google, Meta, Microsoft — have signalled around $630 billion in capex for 2026, a 62% jump on the record $388 billion they spent in 2025. Add Oracle and the figure tops $600 billion across the Big Five, with roughly three-quarters of it tagged for AI infrastructure. Bloomberg has Amazon alone at $200 billion (vs $125 billion last year), Google at $175–185 billion, Meta at $115–135 billion, and Microsoft at $110–120 billion. Crucially, more than 60% of that spend is now flowing into power, cooling and construction — not chips. The bottleneck has visibly moved. Then the supply side. The US interconnection queue stood at roughly 2,060 GW of generation and storage capacity at the end of 2025, per Lawrence Berkeley National Lab — down from a 2,300 GW peak but still more than twice the size of the entire US grid. Median time-to-energisation for a new project is closing in on five years, and for large data center loads in PJM and MISO the wait can stretch towards a decade. Only 13% of capacity that requested interconnection between 2000 and 2019 has actually reached commercial operation; 77% has been withdrawn. Now the transformer crunch — the most boring number in the stack and the most important. Bessemer notes transformer demand rose 119% from 2019 to 2025. Lead times from incumbents (GE Vernova, Siemens Energy, Mitsubishi, Hitachi Energy) have stretched from around one year pre-COVID to as long as five years today. Switchgear lead times now run beyond 60 weeks. High-voltage cable and breaker backlogs are widely flagged as the next constraint. This is where Europe’s industrial champions enter the picture. Siemens Energy closed FY25 with an order backlog of €138 billion — a record — with roughly 85% of FY26 expected revenue already booked. The company is putting €220 million into expanding its Nuremberg transformer plant (350 new jobs) and building its first US transformer factory in Charlotte. Schneider Electric posted Q1 2026 revenue of €9.8 billion, up 11.2% organically, with the energy management segment up nearly 13% and double-digit gains in data center demand; its 2025 acquisition of US liquid-cooling specialist Motivair is paying off. ABB, Hitachi Energy and Prysmian (HVDC cables) all report comparable AI-driven order books. On the policy side, Trump’s 20 April 2026 DPA Section 303 determination covers transformers, transmission lines and conductors, substations, high-voltage circuit breakers, power control electronics, protective relay systems, capacitor banks and electrical core steel. It waives section 303(a)(1)–(a)(6) — meaning DOE can extend purchase commitments, loans and loan guarantees, and direct subsidies without the usual Congressional notification waits. Pillsbury and Akin Gump both flagged the determination as potentially the largest peacetime industrial policy intervention in the US power sector since the 1930s. Finally the European backstop. Bundesnetzagentur’s draft Netzentwicklungsplan 2037/2045 (March 2026) shows SuedLink — the HVDC corridor meant to move North Sea wind to Bavarian industry — now slipping to 2029. Amazon has indefinitely postponed a €7 billion German campus. Google walked from Mittenwalde. Only two German data centers above 100 MW have been connected so far. Germany must triple electricity supply for data centers in twelve years while decommissioning coal and electrifying transport. The maths does not currently work.

Here is the trade DACH-based CIOs and CFOs actually need to internalise: in a world where compute is bottlenecked by copper, steel and 800V DC architecture rather than by silicon, the most structurally advantaged suppliers are European industrials, not Silicon Valley startups. Siemens Energy’s Grid Technologies division — transformers, switchgear, HVDC converters — is essentially sold out through 2027. The €138 billion backlog at FY25 close is not a vanity number; it is a multi-year revenue lock that gives Siemens Energy the pricing power of a near-monopolist on the chokepoint of the entire AI buildout. CEO Christian Bruch told analysts in November that the company is now turning away orders to protect delivery dates for strategic customers — the kind of statement that, in any other industrial cycle, would have triggered an analyst stampede. ABB and Schneider Electric occupy the next ring. ABB’s Electrification business is reporting AI-data-center orders growing well above 30% year-on-year; Schneider’s Q1 2026 print confirmed the same dynamic, with its Galaxy VXL UPS and the Motivair-derived liquid cooling line both supply-constrained. Hitachi Energy — the former ABB Power Grids unit — is the dominant HVDC converter supplier globally and is sold out into 2028. Prysmian and Nexans control most of the world’s high-voltage submarine and underground cable manufacturing capacity; both have order books extending into the 2030s. On the generation side, RWE has identified the Rhineland — converting retiring coal plants in the Aachen–Bonn–Cologne–Düsseldorf quadrant — as Europe’s most attractive corridor for hyperscaler co-location. EnBW signed a 15-year, 100 MW PPA with Google in February 2026 from its He Dreiht offshore wind project. French nuclear, via EDF, is quietly running similar conversations with Microsoft and AWS. The pattern is the same on both sides of the Rhine: the hyperscalers are now negotiating direct, multi-decade offtake with European utilities because waiting for the public grid is no longer a viable timeline. The European industrial stack, much-derided through the 2010s as too slow and too regulated, turns out to be exactly the supplier base the American AI buildout cannot do without.

A humanoid robot is a bipedal machine with arms, hands and a vision system, designed to slot into workspaces built for people — a car-assembly station, a warehouse aisle, a hotel kitchen. For most of the last decade the category was a science-fair exhibit: Honda’s Asimo, Boston Dynamics’ Atlas, a handful of academic prototypes. In 2024 and 2025 that changed. A wave of startups — Figure (now FIG.AI), Apptronik, 1X, AgiBot, China’s Unitree — began shipping units to paying customers, mostly carmakers and logistics firms. On March 20, 2026, Unitree Robotics filed a prospectus with the Shanghai Stock Exchange’s STAR Market seeking ¥4.2 billion (about $610 million). It is the first time a humanoid maker has published an audited income statement at scale. Every DAX40 procurement office that has ever asked the question — what does one of these actually cost — now has a public answer.

On a Friday morning in Hangzhou, Wang Xingxing, the 35-year-old founder of Unitree Robotics, signed off on a 600-page prospectus that landed at the Shanghai Stock Exchange before noon. Twelve days later, on April 1, the China Securities Association sent an inspection team to the company’s Binjiang district headquarters, a randomized on-site audit that has stalled more than a few STAR Market candidates. The CITIC Securities-led underwriting team had filed for ¥4.2 billion (roughly $610 million), targeting a post-money in the ¥30-50 billion range, with grey-market trades implying upside above ¥100 billion. The numbers in the document were what stopped traders mid-coffee. Revenue in 2025: ¥1.708 billion, up 335% year-on-year. Adjusted net profit: ¥600 million, up 674%. Humanoid units shipped: more than 5,500. Humanoids — barely 1.9% of revenue in 2023 — were 51.5% of core revenue in the first nine months of 2025. A category that consultants had been pricing as a 2030 story now had an audited 2025 P&L attached to it. By Sunday the filing had been chewed through twice by the people DAX40 chief technology officers actually read. Tanay Jaipuria of Wing Venture Capital, writing in his weekly note republished in the Sandhill.io VC digest, called it the first filing that “gives a good sense of the current state of the robotics market” — a category that, until last week, had no audited reference point at all. In the same Sandhill issue, Jordan Kretchmer of Outlander VC argued that what physical AI is now missing is not money or compute but a common interface layer — an MCP for the physical world — that would let any vision-language-action model drive any humanoid body the way MCP lets any LLM drive any tool. Put together, the two pieces re-priced the week: open weights from Alibaba’s RynnBrain release in February, open hardware reference points from Unitree’s prospectus, and an open-standards proposal from a US seed-stage VC. The scene that German readers should hold in their head sits 7,500 kilometres from Hangzhou. At Mercedes-Benz’s Digital Factory Campus in Berlin-Marienfelde, a small fleet of Apptronik Apollo robots — backed by a €100 million Mercedes investment — has been moving sub-assemblies between stations since 2025. At BMW’s Leipzig plant, the first European deployment of Figure (now FIG.AI) humanoids went live in December, after Figure 02 completed a ten-month Spartanburg run helping build 30,000 X3s. None of those buyers had, before March 20, a single line of audited financials from a competitor maker. They do now — and it comes from Hangzhou.

The Unitree prospectus is the densest pricing document the humanoid category has produced. Start with the headline: ¥1.708 billion revenue in 2025, up 335% from ¥392 million in 2024. Adjusted net profit was ¥600 million, up 674%. Gross margin across the company hit 59.5% in the first three quarters of 2025, up from the mid-40s in 2022-2023. The mix shift is what makes the document interesting. In 2023, humanoids contributed 1.9% of revenue and quadrupeds (the Go2 robot dog and its industrial cousin B2) did almost everything else. Through three quarters of 2025, humanoids were 51.5% of core revenue, quadrupeds 42.3%. Average humanoid selling price collapsed from roughly ¥593,400 in 2023 (~$85,000) to ¥167,600 in 2025 (~$25,000) — a 72% cut in two years, made possible because Unitree builds its own motors and actuators and pays only 14-18% of cost of goods to external component suppliers. Proceeds are earmarked for AI model research, new product development and manufacturing capacity. Wang Xingxing retains a 23.8% direct stake and 69% of voting rights through a dual-class structure that, by Chinese standards, is unusually founder-friendly. Reported pre-IPO valuation ranges put the company at $3-7 billion; secondary-market quotes have implied above ¥100 billion (~$14 billion) on listing. Benchmark this against the private comparables. Figure / FIG.AI was last reported at a $39.5 billion valuation on a 2025 round; 1X closed a $100 million Series B in early 2024 and has since added growth capital around home and warehouse pilots; Apptronik has taken Mercedes money at undisclosed terms; AgiBot, Unitree’s closest mainland rival, is rumored to be raising at a $7 billion mark; Boston Dynamics’ Atlas remains inside Hyundai. Germany’s Neura Robotics — the only European pure-play at scale — closed a Tether-led Series C at a €4 billion valuation on March 26, 2026, six days after Unitree filed. UBTECH, the other Chinese listed humanoid maker, lost ¥700 million in 2025 against Unitree’s ¥600 million profit on similar revenue. That single comparison — same country, same category, opposite signs on the bottom line — is what makes Unitree’s document the new floor on the conversation. The caveat that every DAX40 buyer will read first: of humanoid revenue in 2025, 73.6% came from research and education customers — universities, labs, R&D pilots. The Robot Report’s analysis of the filing put it sharply: a real hardware business, an early humanoid commercial case. The 5,500-unit shipment number includes G1 desktop-class humanoids selling at ¥99,000 to professors. The industrial-grade H1 ships in volumes a fraction of that. BYD’s 2018-2023 electric-vehicle ramp — from a domestic curiosity to a global cost leader — is the template Wang is selling. Whether Unitree’s 2026 R&D-heavy revenue mix can graduate into recurring industrial contracts is the question the prospectus does not answer.

Mercedes-Benz is the deepest in. Berlin-Marienfelde and the Kecskemét plant in Hungary host single-digit fleets of Apptronik Apollo units running intralogistics; Mercedes has put more than €100 million into Apptronik and signed the only published commercial agreement, not a research MoU. BMW followed with a different bet. After Figure 02 completed a ten-month, ten-hour-per-shift production run at Spartanburg in 2025 — contributing to 30,000 X3s — BMW announced on February 27, 2026 that AEON humanoids would enter Plant Leipzig, with a December 2025 test deployment, an April 2026 expansion and the pilot phase proper from summer 2026. The Leipzig site has been designated BMW’s Center of Competence for Physical AI in Production. Volkswagen Group is less public. Group software arm CARIAD has worked with Wandelbots (Dresden) and NVIDIA Isaac Sim on robotics deployment tooling at Volkswagen Sachsen; the Wolfsburg pilot floor has hosted 1X NEO units in closed evaluation, though no commercial agreement has been disclosed. Audi’s parent has separately watched UBTECH’s Walker S Lite enter the FAW-Volkswagen plant in Qingdao — a useful, if uncomfortable, control case for what Chinese humanoids on a European-brand line actually look like. Bosch is hedged through Neura Robotics, the Metzingen-based maker of the 4NE1 humanoid, in a joint manufacturing and AI partnership signed in 2025; Tether’s €1 billion-plus Series C in March 2026 valued Neura at €4 billion, the largest European pure-play number to date. KUKA (Augsburg, owned by Midea since 2016) has signaled a humanoid roadmap but no commercial product. Festo continues to ride its bionic-design pedigree without a bipedal commercial play. Siemens supplies the digital-twin software (Xcelerator, Industrial Edge) that most of these pilots run on, a quiet but defensible position. Stellantis and ZF have run smaller, less publicized trials. The Mittelstand integrator base — Pilz, B&R Industrial Automation, the broader VDMA Robotics group — is watching procurement signals and waiting.