Daily 2026 · 05 · 31
Tech Economic

SoftBank books up to €75 billion to build 5 gigawatts of AI data centers across northern France

Masayoshi Son commits an initial €45 billion to deliver 3.1 gigawatts of capacity in Dunkirk, Bosquel and Bouchain by 2031, with sites built and operated alongside Schneider Electric, EDF and SB Energy. The package, unveiled at Emmanuel Macron's Choose France summit, is the largest European AI infrastructure investment any single firm has ever announced. It lands days after SoftBank disclosed a 9.2-gigawatt natural-gas-backed campus in Ohio.

SoftBank Group · 2026-05-30 · Source ↗ ·

SoftBank Group disclosed plans to invest up to €75 billion (about $87 billion at current rates) to build 5 gigawatts of AI data center capacity in France, with an initial €45 billion committed to deliver 3.1 gigawatts of capacity in the Hauts-de-France region by 2031. The first three sites sit in Dunkirk (the Loon-Plage industrial zone at the port), Bosquel near Amiens, and Bouchain on the site of an existing EDF gas turbine plant. Masayoshi Son announced the commitment at the 2026 edition of Emmanuel Macron's Choose France summit at the Château de Versailles, where French ministers court foreign capital one weekend a year.

The deal vaults France into the second tier of national AI infrastructure plays behind the United States. It is the largest single AI infrastructure commitment any company has ever directed to a European country, and SoftBank itself describes the package as its largest AI infrastructure investment outside the U.S. SoftBank Energy will originate power supply, EDF will provide grid interconnection and low-carbon generation, and Schneider Electric will partner on the build-out and operate a power-module integration facility at the Port of Dunkirk that will sit next to a SoftBank-run enclosure manufacturing line. French Economic Minister Roland Lescure framed the announcement as evidence that "President Macron's ambition to position France as a leading destination all along the AI value chain" is taking hold.

What gets built, where, and when

The plan covers three buckets. Phase one is the 3.1-gigawatt initial buildout in the three named sites; SoftBank says construction begins inside 18 months at the Dunkirk port-adjacent campus and at Bouchain, with Bosquel following. Phase two takes the program to 5 gigawatts through additional sites yet to be named. The third bucket is the industrial cluster at the Port of Dunkirk, where SoftBank and Schneider Electric will assemble enclosures, power distribution modules and cooling integration for the program and, the companies say, eventually for export to other European data center developers. SoftBank's framing leans on the jobs counts a Choose France announcement is expected to deliver: thousands of high-skill construction, operations and manufacturing roles, alongside regional university partnerships.

The 3.1-gigawatt number invites a direct comparison with the U.S. Stargate program, where SoftBank, OpenAI and Oracle are the lead participants and individual campuses are sized at multiples of one gigawatt. SoftBank, an investor in and customer of OpenAI, did not name an anchor tenant for the French sites. French and EU officials have repeatedly named Mistral as a national champion in need of compute, and the choice of Hauts-de-France keeps the buildout inside the European Union for sovereignty and AI Act compliance purposes. The Dunkirk port allocation also keeps the option of grid-edge generation alive: the existing French nuclear baseload routes through Bouchain, and the Loon-Plage zone has direct access to the offshore wind buildout in the North Sea.

"AI is entering a new era, and the countries that build the infrastructure for this transformation will shape the future."— Masayoshi Son, Chairman and CEO, SoftBank Group

The package against SoftBank's broader compute ledger

The France commitment lands two weeks after SoftBank confirmed a 9.2-gigawatt natural-gas-backed campus in Ohio, and it sits alongside the Stargate commitments of the past 18 months. Stack the new French capacity against the SoftBank-allocated portions of Stargate and the company's effective pipeline is now north of 20 gigawatts, an order of magnitude above any AI infrastructure portfolio that existed three years ago. The capital comes from SoftBank Group's balance sheet, with project-level co-investors expected for the French sites and the Hauts-de-France region likely to provide tax abatements and accelerated permitting in exchange for the jobs.

The contrast with U.S. data center politics is the subtext of the day. TechCrunch noted that SoftBank's separate Ohio commitment had drawn opposition over the 9.2-gigawatt gas plant required to power it, and several Texas and Virginia campuses faced ratepayer pushback this spring. France, with its EDF-led nuclear baseload and an industrial policy that treats heavy electricity loads as a feature rather than a problem, offers a backstop the American grid does not. The French sites are designed to be powered almost entirely without new fossil generation, an explicit pitch SoftBank used in the announcement and one the company has not been able to make about its U.S. campuses.

"This investment demonstrates President Emmanuel Macron's ambition to position France as a leading destination all along the AI value chain."— Roland Lescure, French Minister of Economy

Why the French location matters for the AI-policy fight

The buildout is the largest AI infrastructure investment ever directed to the European Union. It immediately changes the political math around the AI Act and the Commission's debates over compute thresholds, sovereign AI, and competitive cloud markets. Stargate sits in the United States. The major Chinese campuses sit in Inner Mongolia and Guizhou. Until today, the EU had no comparable single-country anchor; today, France does. Mistral, the Commission, and France's broader Choose France agenda will all carry the SoftBank commitment forward as evidence that the bloc has secured a meaningful share of the global AI infrastructure capex without using sovereign capital to do so. The investment will also raise the bar for Germany, the U.K. and the Nordics, all of which have separately courted similar capital with smaller offers.

Risks remain. The 2031 delivery target sits inside two French presidential cycles, and Hauts-de-France's grid build-out depends on EDF capex that has slipped repeatedly. SoftBank's own cash flow profile is leveraged against the same Vision Fund investments that have produced uneven returns; the company will have to find project finance partners for at least part of the €75 billion. And the absence of a named anchor tenant means the build risk falls on SoftBank if Mistral or another European champion fails to absorb the capacity at the planned rates. None of that disqualifies the announcement; all of it sets the bar for what the next 24 months of execution have to look like.

Big picture

SoftBank's €75 billion French data center package is the day's anchor and the largest AI infrastructure commitment any single firm has ever made in Europe. Phase one delivers 3.1 gigawatts in Dunkirk, Bosquel and Bouchain by 2031; the ultimate target is 5 gigawatts, sitting alongside Stargate in the U.S. and SoftBank's separate 9.2-gigawatt Ohio campus. The announcement gives the EU its first AI infrastructure anchor of comparable scale to the U.S. and Chinese hubs.

The day's political story is parallel. The New York Times reported the two largest AI labs are now funding rival super PAC operations heading into the 2026 midterms: OpenAI president Greg Brockman has put $25 million into Leading the Future and $25 million into Trump's MAGA Inc, while Anthropic seeded a corporate AnthroPAC, gave $20 million to Public First, and now backs candidates in the Bores–Wiener regulatory wing. The Wall Street Journal's profile of Brockman the same day described him as the OpenAI executive most visible to the Trump administration.

Three currents move under the headlines. Enterprise tokenmaxxing is collapsing on contact with FY budgets: Uber burned its 2026 token allocation in four months on Claude Code, Amazon killed its Kiro Rank internal leaderboard after employees gamed it, Microsoft cut Claude Code subscriptions in several divisions, and Salesforce now expects a $300 million Anthropic bill this year. MiniMax filed for a Shanghai IPO on $300 million of ARR, doubled in two months after its M2.7 release. And memory chips moved further out of spot-market reach as Samsung, SK hynix and Micron each crossed $1 trillion in market value last week, with HBM allocation now locked years out under multi-year contracts.

More today

12 stories · 2026-05-30 — 2026-05-31
PolicyEconomic

Anthropic and OpenAI back rival super PACs heading into the 2026 midterms

The New York Times · 2026-05-30 · Source ↗ ·

OpenAI's executives and venture allies are seeding Leading the Future to fight new state AI rules; Anthropic's employees, a fresh AnthroPAC, and a $20 million grant to Public First push the other way.

The Times catalogued the AI industry's first full midterm cycle as a two-camp fight. OpenAI president Greg Brockman and his wife Anna gave a combined $50 million split between Leading the Future, a nominally bipartisan AI super PAC, and MAGA Inc, Donald Trump's own super PAC. Marc Andreessen and Ben Horowitz seeded Leading the Future alongside Brockman, and the group is now spending in House and Senate primaries against candidates the AI-accelerationist wing considers regulation-friendly. Leading the Future's stated agenda is a single federal AI framework that preempts state laws.

Anthropic is on the other side, in three vehicles. AnthroPAC, a traditional corporate political action committee, raised $119,019 from employees in its first quarter, with individual contributions capped at $5,000. Anthropic the company donated $20 million to Public First, a political nonprofit whose affiliated super PACs (the bipartisan Public First PAC, the Democratic-aligned Jobs and Democracy PAC, and Defending Our Values PAC for Republicans) are run by former Reps. Brad Carson and Chris Stewart. And Anthropic employees have moved $880,000 across 302 hard-money donations this cycle, an average of more than $12,500 per donor.

The race-by-race picture

The two clearest beneficiaries of Anthropic-aligned giving are state-level regulatory champions: California state senator Scott Wiener, author of SB 53, has received more than $110,000 in Anthropic-employee hard money, and New York assemblymember Alex Bores, author of the RAISE Act, has pulled in roughly $186,000. Both bills set frontier-model safety reporting requirements, and both are exactly the kind of state action Leading the Future's federal-preemption pitch is designed to neutralize.

"The campaign is the one that is able to put out the message it controls and pay the people and pay the staff."— Brendan Steinhauser, political strategist, on the strategic value of hard-money donations

The split tracks the labs' public posture. OpenAI's policy team has spent the past year arguing that fragmented state rules will hand China the next decade of AI capability; Anthropic's safety leadership has lined up behind sector-specific reporting and pre-deployment evaluations, and CEO Dario Amodei testified in favor of California's SB 53 last year. The midterms will be the first test of how well that posture translates into legislative seats, and the first cycle in which AI money is large enough to read independently against crypto and traditional tech political spend.

EconomicPolicy

Greg Brockman becomes OpenAI's most visible executive to the Trump administration

The Wall Street Journal · 2026-05-30 · Source ↗ ·

The OpenAI co-founder, long content to stay behind the product, now runs the company's product organization and acts as its ambassador to the White House after $50 million in personal political giving.

The Journal's Sunday profile centered on Greg Brockman's pivot from engineer-in-residence to OpenAI's external face. Brockman, the co-founder who stepped away briefly in 2024 and returned to oversee product, now runs the company's product organization day-to-day, and the Journal reports he is the executive most often in the room when OpenAI talks to the Trump administration. The profile follows his September 2025 decision to put $25 million into MAGA Inc and another $25 million into Leading the Future, the AI-industry super PAC he and his wife Anna helped seed.

OpenAI staff have not taken the political turn quietly. The Journal cites internal disagreement over the donations and a cancellation campaign that surfaced on X after the giving was reported. Brockman has defended the spending as backing "policies that advance American innovation and constructive dialogue between government and the technology sector," language that mirrors OpenAI's official position that federal AI rules should preempt state legislation. The Journal frames the new visibility against Sam Altman's reduced public profile in 2026 and the Elon Musk lawsuit against OpenAI, in which Brockman is a named defendant and which entered trial testimony in late April.

Brockman's portfolio inside OpenAI

Inside OpenAI, Brockman owns product, model launch coordination, and the agent stack. ChatGPT product features now route through his organization, and the Codex Computer Use rollout for Windows shipped under his lead. Externally, he is the executive who has met with the White House CTO's office and the OSTP team on Stargate progress and on Leading the Future's preemption push. Allies inside the company describe him as the most product-fluent communicator OpenAI has; critics describe the political spending as an unforced own-goal that has put a target on the company at exactly the moment Anthropic moved past it on revenue.

"Our donations express support for policies that advance American innovation and constructive dialogue between government and the technology sector."— Greg Brockman, President, OpenAI

The Journal's underlying claim is that Brockman is now the executive whose decisions move OpenAI fastest. Altman remains CEO and remains the face when investors are in the room; Brockman is the one customers and regulators talk to. The reframing matters because the company has spent two years rotating senior people through outward-facing roles, and the profile is the clearest signal yet that the rotation has stopped at Brockman.

ApplicationTrend

"Tokenmaxxing" collapses as Uber, Amazon and Microsoft ration AI usage

Wall Street Journal / Fortune / FT · 2026-05-30 · Source ↗ ·

Uber burned its full 2026 token budget in four months; Amazon shut down the Kiro Rank internal leaderboard; Microsoft canceled Claude Code seats in product divisions; Salesforce expects a $300 million Anthropic bill this year.

The WSJ piece this weekend is the clearest cross-company picture yet of the productivity-versus-cost reckoning AI customers are running. Uber chief operating officer Andrew Macdonald told a recent podcast the company exhausted its entire 2026 token budget in the first four months of the year, with Claude Code the largest line item. Inside Amazon, employees were observed spinning up agents to perform unnecessary work just to keep their numbers on Kiro Rank, the internal leaderboard that tracked individual AI tool use; Amazon scrapped the leaderboard. Microsoft canceled Claude Code subscriptions in several product divisions after similar overspend, per The Verge's reporting cited in the piece, while Meta took down an informal token leaderboard staff had built.

The headline number on the other side is Salesforce. CEO Marc Benioff has said publicly that Salesforce's Anthropic bill will run about $300 million in 2026. Across Fortune's coverage, executives describe the same pattern: aggressive token consumption was rewarded internally as a proxy for engagement, the reward structure produced Goodhart's Law in textbook form, and the bills came in at multiples of the original plan with no clear company-wide productivity number to point at.

What changes from here

Three things follow from the reset. First, the inference-revenue base case for Anthropic and OpenAI has to be re-underwritten: if the largest customers ration usage and kill the leaderboards, the per-seat token velocity that anchored 2026 forecasts compresses. Second, the internal-tools market for "AI cost management" suddenly has runway, and the early movers (LangSmith pricing analytics, Helicone, OpenRouter's gateway billing) will get pulled into procurement faster than the budget cycle would have implied. Third, the agentic-coding category will get reframed: the question is no longer "how many tokens per dev per day" but "how many merged PRs per token spent," which favors models with better tool-use efficiency and worse-looking raw consumption.

"The reward structure created exactly the behaviour the metric was supposed to detect. Big Tech is now scrapping leaderboards that lack substance."— Fortune, on the unwinding of internal tokenmaxxing programs

The reset does not change the model providers' top line for 2026 — pre-paid contracts and committed-use agreements bake in the revenue — but it changes the slope going into the next renewal cycle. Anthropic and OpenAI have both been pricing on the assumption that enterprise per-seat consumption keeps doubling every quarter; the WSJ and FT reporting this weekend makes that assumption harder to defend in the second half.

TechEconomic

MiniMax files for a Shanghai IPO on $300 million ARR, doubled in two months

Bloomberg · 2026-05-30 · Source ↗ ·

The Shanghai lab files for a mainland listing five months after a Hong Kong float, citing more than 1 million enterprise users and revenue more than doubled since the March release of its M2.7 model.

MiniMax Group filed for a domestic A-share IPO in China, with the regulatory filing disclosed over the weekend and reported first by Bloomberg. The Shanghai-based lab listed on the Hong Kong Stock Exchange in January and is now pursuing the dual listing to fund further model development and to give onshore Chinese investors a direct vehicle. The filing follows last week's revenue disclosure that annualized revenue had more than doubled in two months to at least $300 million, an outcome MiniMax attributed to its mid-March M2.7 model release and the resulting wave of enterprise adoption.

Enterprise users of MiniMax's API and managed agent services now exceed one million, up roughly fivefold over six months. The company is the largest Chinese AI lab outside of DeepSeek and Alibaba's Qwen team to publish revenue at this level. The mainland listing is meant to address two constraints at once: capital for the next model generation, and the political capital that comes with being listed on a Chinese exchange when Beijing is pushing for a domestic AI champion in the wake of the DeepSeek breakout. MiniMax's M2.7 reasoning model has been the standout performer on agentic-coding benchmarks among Chinese open-weight contenders, and the company has shifted toward closed weights for its newest model class as it has scaled commercial revenue.

"M2.7 changed the demand curve overnight. The IPO is a function of needing capital to keep up with the inference build-out the model has triggered."— Industry observer quoted in Bloomberg coverage

The filing also opens the question of how Beijing handles a major AI lab going public during a period when the U.S. government's AI export-control framework is being rewritten under the Trump executive order. MiniMax's customer base is overwhelmingly Chinese, but the company has begun signing Southeast Asian and Middle Eastern enterprise contracts; an A-share float will subject those to additional disclosure. The timing also implicitly responds to DeepSeek's continued resistance to a public listing, leaving MiniMax to define what an investable Chinese AI lab looks like.

Economic

Anthropic cuts its unauthorized secondary-market platform list in half

Bloomberg · 2026-05-30 · Source ↗ ·

After pushback from secondary-market investors, the lab shrinks its flagged-platforms list from eight to four, a small but read-through move on how late-stage AI shares change hands at $965 billion valuation.

Anthropic reduced from eight to four the number of secondary-market platforms it had publicly flagged as unauthorized to facilitate trades in its shares, Bloomberg reported. The original list, issued in the days after the $65 billion Series H closed at a $965 billion post-money valuation on May 28, was meant to put platforms and brokers on notice that any sale or pledge of Anthropic equity without the company's consent would not be honored. Investors and brokers on the original list pushed back, arguing the wider listing improperly tarred legitimate transfer processes that had been pre-cleared.

The compromise list keeps four platforms named — Anthropic did not disclose which ones it removed, citing ongoing discussions — and the company is now in direct contact with the rest of the secondary market to define a permitted-transfer process before the IPO. The Series H raised the stakes considerably. Late-stage AI shares are now traded in size on multiple platforms, and Anthropic's $965 billion mark put the company in the same secondary-market liquidity tier as Stripe and SpaceX. The unauthorized-platform list is a tool the company has used before, but the public version this week is the most aggressive enforcement action a frontier lab has taken on its own cap table.

The smaller list still serves its purpose: brokers attempting to facilitate Anthropic transfers through any of the four named venues now face a documented disqualifying notice. The signal to the broader secondary market is that the company is willing to revisit the list as needed, both to relax it when satisfied and to expand it when not. With an IPO confidentially filed in April and a likely 2027 listing, the cap table needs to clean up well in advance, and the secondary platforms know it.

ApplicationTech

Meta plans an AI pendant for 2027, new smart glasses next month, and a Wearables for Work unit

The Information · 2026-05-30 · Source ↗ ·

Reality Labs' internal roadmap leaks: new smart glasses ship in June, an always-on AI pendant enters testing in 2027, and a new "Wearables for Work" unit will sell directly to enterprises.

The Information published an internal Meta memo that lays out Reality Labs' wearables roadmap for the next 18 months. The most surprising line item is an AI pendant that the team intends to begin internal testing in 2027 — Meta's first non-glasses Reality Labs hardware product. The pendant would clip to a shirt or hang from a chain and host always-on microphones, cameras and an on-device language model that streams to cloud Llama for heavy lifting. The memo, by Reality Labs head Andrew Bosworth, acknowledges internal skepticism about the form factor, including from designers who worry about the social acceptability of an always-listening object worn on the chest.

The shorter-horizon items are firmer. New smart glasses ship next month, with two new SKUs, expanded language support and a longer-battery athletic frame. A separate enterprise unit, "Wearables for Work," will be stood up later this year to sell directly to corporations. The unit's positioning matches what Microsoft, Google and OpenAI have done with forward-deployed-engineer programs: embed product and engineering into the largest customers rather than selling at arm's length. Bosworth's pitch is that Meta's wearables generate more useful telemetry inside a real workflow than any consumer product, and that the enterprise team will close the loop.

Why this matters for the AI hardware fight

The pendant filing is the first credible non-Humane attempt at the always-on AI wearable since the Humane Pin's failure. The form factor matters less than the underlying claim: that Meta now believes the on-device LLM stack — likely a Llama 4 distillation — is mature enough to ship into a body-worn product that customers will pay for. That is a different bet than the smart-glasses bet, which is principally a display problem. The Wearables for Work unit, meanwhile, looks like Meta's answer to Microsoft's Copilot devices push and to Apple's still-unannounced Vision Pro enterprise pivot.

The internal skepticism is real and reported. The Information cited at least two product leaders who pushed back on the pendant during the memo's distribution, arguing that consumer comfort with a worn microphone has not been demonstrated and that the Llama-on-device throughput estimates are aggressive. Bosworth's framing is that the company has the runway to test the form factor and walk away if it does not work; the public-roadmap version is the smart-glasses cadence, which is on track and incrementally improving.

ApplicationTech

Wired runs Gemini Spark through a real week, finds the agent useful and the relationship classifier blunt

Wired · 2026-05-30 · Source ↗ ·

Google's $100-a-month always-on agent now plans events from inbox plus calendar context, but described a tester's live-in partner as a "close friend," underscoring the relationship-classifier gap Wired flagged in I/O previews.

Wired published the first non-Google long-form hands-on of Gemini Spark, the always-on personal AI agent that began rolling out to U.S. Google AI Ultra subscribers the week of May 26 at the newly reduced $100/month tier. Spark runs on a dedicated Google Cloud VM that persists across the user's devices and continues working when the phone is locked or the laptop is shut. The agent is built on Gemini 3.5 Flash with Google's Antigravity 2.0 harness and it reads from connected Gmail, Calendar, Docs, Chrome, and a growing list of third-party connectors.

The reviewer's headline finding: Spark planned a 30-person birthday party end to end from inbox context and a few short prompts, including venue research, vendor outreach, and a draft of follow-up emails for sign-off. The relationship-context limitation surfaced when the agent, asked to draft a message to the reviewer's live-in partner about logistics, addressed them as a "close friend" and did not adjust tone when corrected mid-conversation. Wired argues the gap will sort itself out as memory accumulates but that the cold-start experience is currently the weak link.

Where Spark sits against Anthropic's Cowork and OpenAI's Codex Cloud

Spark is the second consumer-facing always-on agent on the market after Anthropic's Claude Cowork, and the first to combine cloud persistence with deep first-party Google productivity integration. The Wired piece quotes pricing alongside ChatGPT Pro's $200/month and Anthropic's $100/month Claude Max with Cowork; the AI Ultra tier was cut from $250 to $100 specifically to compete in this band. Google's pitch is that the Gmail/Calendar/Drive data graph is what gives Spark its leg up on planning tasks; Anthropic's pitch is that Cowork's VM model and integrated browser is better for execution-heavy work.

The 24/7 cloud-VM model also opens a security surface the Anthropic engineering team wrote about earlier this week. Persistent agents accumulate context that survives between sessions, which creates new failure modes around memory poisoning and credential exfiltration. Wired notes Google has not published a comparable containment architecture for Spark, and that the only safeguard the user can directly inspect is the per-connector permission grid.

TechEconomic

Memory chip makers leverage AI demand into multi-year locked contracts

The Wall Street Journal · 2026-05-30 · Source ↗ ·

SK hynix, Samsung and Micron crossed $1 trillion each last week. The WSJ reports there is no spot market left for HBM in 2026 and that 2027–2028 allocations are now signed in firm contracts to a handful of AI buyers.

The WSJ's weekend piece picks up where the trillion-dollar market-cap milestones left off. As of last week, Samsung, SK hynix and Micron each carry market values above $1 trillion, a combined $3+ trillion oligopoly that did not exist 18 months ago. The structural story behind the prints is that the high-bandwidth memory the largest AI buyers need is no longer available on the spot market — every wafer that can be allocated to HBM has been forward-sold under multi-year contracts. SK hynix retains the largest share of HBM revenue at roughly 57%, with Samsung at 22% and Micron at 21% as of Q3 2025; all three have shifted material DRAM wafer capacity into HBM production.

The contracts side has the most operational read-through. The October 2025 letter of intent under which Samsung Electronics and SK hynix agreed to supply OpenAI with up to 900,000 DRAM wafers per month for Stargate is now public; Nvidia has the majority of SK hynix's HBM output through 2026 under separate multi-year agreements; and the second-tier buyers — Anthropic via its hyperscaler partners, the Chinese labs, the European sovereign clouds — are now competing for a residual allocation that is smaller than the cumulative demand the same buyers projected six months ago.

The downstream effects

Three follow on. First, the cost of training compute has stopped declining in the way it had through 2024–2025: cluster-level training cost-per-token is now flat or rising because HBM is the binding constraint. Second, consumer DRAM and DDR5 supply is being squeezed; smartphone bills of materials are up 14% year-over-year because the wafer mix shifted to HBM. Third, the strategic-investor allocations on the Anthropic Series H — Samsung, SK hynix and Micron each took meaningful pieces of the round — read very differently in light of the contract picture. Memory supply is now a fundraising primitive, not just a procurement line.

The WSJ's framing — that memory chips are now "more valuable than oil" — overstates the comparison on absolute dollar value, but it captures the structural fact that HBM allocation is the binding constraint for AI growth through at least 2027. The labs that lock supply now ship the next two model generations; the labs that do not are betting on capacity that does not yet exist.

EconomicTrend

VC robotics and physical-AI funding hit $26 billion in 2025, up from $4.2 billion in 2019

The Wall Street Journal · 2026-05-30 · Source ↗ ·

Kate Clark catalogs the venture turn into hardware: $26 billion into robotics and physical AI in 2025, more than double 2024, as VCs hedge against the model providers' direct path into application software.

The Wall Street Journal's Kate Clark put a number on the shift the entire venture industry has been quietly executing. Global VC investment in robotics and physical AI reached roughly $26 billion in 2025, up from $4.2 billion in 2019 and more than double 2024's $13.7 billion, with 1,000-plus disclosed deals. The headline number puts hardware on a faster trajectory than the AI applications category that funded the 2023–2024 cycle, and the reason VCs gave Clark on background is consistent: the model providers' aggressive entry into application software (Anthropic into agentic coding, OpenAI into enterprise deployment) has made horizontal software bets riskier, and hardware is harder to disrupt with a model release.

The deal mix Clark describes is dominated by humanoid robotics (Figure, 1X, Apptronik), warehouse autonomy (Symbotic, Locus, Covariant), foundation-model robotics labs (Physical Intelligence, Skild, Pi Labs), and defense-adjacent integration plays (Anduril's robotics extensions, Shield AI, Picogrid). The defense-tech share specifically has tripled year-over-year; Picogrid's $45 million Series A this week, led by Bessemer, is the most recent illustration. Drone-related deals contributed a 139% surge in defense-robotics funding tracked separately by other industry data.

The hedge thesis

Clark's reporting frames the shift as a hedge against software-AI consolidation. If the largest model providers can ship vertical applications on top of their own model platforms — Anthropic's Code, OpenAI's enterprise apps, Google's Spark agent — then the addressable market for independent application software shrinks. Hardware investments require physical supply chains, manufacturing partnerships, and on-the-ground deployment teams that the model labs have so far chosen not to build. The bet is that physical AI is the right place to put dollars during a period when software AI consolidates around three or four vendors.

The risk is that physical AI also consolidates around the model providers eventually. NVIDIA's robotics platform partnerships with BlackBerry QNX, Toyota's Woven manufacturing program, and Google's Spot integration with DeepMind's Gemini Robotics 1.6 all point to the same end state: the model lab provides the brain, the platform vendor provides the runtime, and the startup ends up in a similar squeeze to the application-software wave. For now, the dollars are flowing in regardless, and the headline number is what investors will quote for the next two quarters.

Tech

Project Jupiter switches to fuel cells, lands a 2.8-GW OpenAI campus in New Mexico

CleanTechnica · 2026-05-30 · Source ↗ ·

BorderPlex Digital Assets revises the Santa Teresa site to fuel cells, cuts water use 50× and NOx 92% versus a gas-fired alternative, but the campus would still emit 10.1 million tons of greenhouse gases a year.

BorderPlex Digital Assets disclosed the revised power architecture for Project Jupiter, the 2.8-gigawatt data center campus it is building in Santa Teresa, New Mexico, on behalf of an Oracle-OpenAI tenant. The site, near the El Paso, Texas border, was originally planned with natural-gas-fired generation. The April 2026 revision replaces it with on-site fuel cells, a configuration that cuts nitrogen-oxide emissions by approximately 92% and water consumption to 20,000 gallons per day, down from roughly 1 million gallons per day under the original gas plan.

The fuel-cell version is the largest behind-the-meter fuel-cell data center on record by a wide margin. Construction started in fall 2025 on the gas plan; the fuel-cell substitution adds capital cost but reduces local environmental opposition and the regulatory exposure of the original air-permit application. The campus will host OpenAI inference and training capacity under the Oracle relationship that anchors the southern portion of the Stargate program.

What the fuel cells don't change

The Center for Biological Diversity flagged the underlying emissions number, which the fuel cells reduce but do not eliminate. Project Jupiter at 2.8 GW would emit approximately 10.1 million tons of greenhouse gases per year — "more than the greenhouse gas emissions from Albuquerque, Las Cruces, and Santa Fe combined," per a Center attorney quoted in the article. The fuel cells improve the NOx and water profile substantially; they do not turn the campus carbon-neutral, and the CleanTechnica piece is explicit that the underlying load is the constraint.

The Project Jupiter announcement lands the same week as SoftBank's 9.2-GW Ohio campus, the Texas grid-interconnection queue's latest revision, and Anthropic's xAI Colossus contract. In aggregate, U.S. AI campus capacity under construction or in late-stage permitting now exceeds 60 gigawatts, and the New Mexico site is a representative case for how the politics of siting are evolving: fuel cells, batteries, and small modular reactor pre-orders are becoming the consolation prizes that allow new gas peakers to be dropped from the project envelope.

ApplicationPolicy

SpaceX wins a $4.16 billion Golden Dome contract to build the space-based tracking layer

Bloomberg · 2026-05-30 · Source ↗ ·

The Space Force awards SpaceX the contract to build, launch and operate a satellite tracking network for the Golden Dome air defense initiative, the largest single defense-AI award of 2026 to date.

The U.S. Space Force awarded SpaceX a $4.16 billion contract to build the space-based tracking layer of the Golden Dome defensive shield, the homeland missile and aircraft defense program announced by President Trump in his first 90 days. The contract covers construction, launch and operation of a satellite constellation that tracks foreign aircraft, cruise missiles, and ballistic missiles in flight, feeding the data into a fused targeting system that links to ground-based and Aegis-class naval interceptors. SpaceX wins the award alongside earlier contracts to Lockheed Martin and Northrop Grumman for separate Golden Dome layers.

The AI angle is the targeting and tracking software. The Space Force's program-of-record specifications call for an onboard ML pipeline running on each satellite, plus a ground-based correlation system that fuses the constellation's sensor data with terrestrial radars and the existing Space-Based Infrared System. Defense industry sources cited by Bloomberg describe the software stack as the largest defense-AI application award of the Trump administration to date, and one that is expected to anchor the procurement of specialized AI hardware (radiation-hardened accelerators, satellite-grade inference chips) over the next 36 months.

The political and industrial context

The award also reads against the political context. SpaceX has been the largest single beneficiary of Trump-administration defense contracts in the first six months, and the Golden Dome award arrives the same week as the Anthropic-OpenAI super PAC story and the Greg Brockman profile. Defense-aligned AI primes — Anduril, Palantir, Shield AI — have been positioning to compete for the software-fusion layer, and the next several Golden Dome awards over the summer will define which AI vendors capture the targeting stack. The European parallel program is reportedly under early design; the EU's own missile defense effort is expected to lean on Mistral and a French defense prime team for the AI layer.

TechSafety

Simon Willison surfaces Anthropic's containment architecture, including a 24-of-25 phishing failure mode

Simon Willison's Weblog · 2026-05-30 · Source ↗ ·

A Saturday post pulls Anthropic's recently published engineering documentation into wider view: gVisor for claude.ai, Seatbelt and Bubblewrap for Claude Code, a full VM for Cowork, and an internal phishing exercise that succeeded 24 times out of 25.

Simon Willison's weekend post highlighted Anthropic's "How we contain Claude across products" engineering documentation, putting the more technical disclosures in front of a wider audience. Anthropic published the document earlier this week with credits to Max McGuinness, Mikaela Grace, Jiri De Jonghe, Jake Eaton and Abel Ribbink. The post lays out the containment architecture across the company's three agentic products: claude.ai runs code server-side inside gVisor containers on ephemeral filesystems; Claude Code runs on user machines with Seatbelt on macOS and Bubblewrap on Linux providing the OS-level sandbox; Claude Cowork runs a full virtual machine using Apple's Virtualization framework on macOS and Hyper-V Container Services on Windows.

The two findings Willison emphasizes are the ones that will get cited in product review. First, an internal phishing exercise against Claude Code's user-as-injection-vector surface — where an attacker crafts a prompt that asks the user to ask Claude to do something malicious — succeeded 24 of 25 times. The model layer cannot reliably distinguish legitimate user intent from social engineering at the prompt level; only environmental controls (filesystem boundaries, egress restrictions) prevent damage. Second, Claude Code users were observed to approve roughly 93% of permission prompts pre-sandbox rollout, the textbook approval-fatigue failure mode; adding the OS sandbox cut the prompt rate by 84% and, by the company's measurement, materially improved the actual safety of execution.

What the post says about persistent-agent threats

The Anthropic team also documented a Cowork-specific vulnerability around domain-allowlist exfiltration: a malicious workspace file used attacker-controlled credentials to upload data through api.anthropic.com, a destination the allowlist permitted by domain. The fix was a man-in-the-middle proxy inside the VM that only accepts VM-provisioned session tokens. That is the kind of disclosure most labs do not publish, and Willison's frustration with the broader industry — that sandboxing products "are rarely thoroughly documented" — runs through the post.

"A complaint I often have about sandboxing products is that they are rarely thoroughly documented, and in the absence of detailed documentation it's hard to know how much I can trust them."— Simon Willison

The disclosure is timely given the rollout of Google Spark this week and the broader push into persistent always-on agents. The threat surface Anthropic describes — persistent memory poisoning, multi-agent trust escalation, the gap between scoped session tokens and cross-platform identity — applies to every product in the category. The headline conclusion is that environmental defense is doing the load-bearing work, and the model layer is not going to fix prompt injection on its own.