Google has a research project called Suncatcher. The idea is to stop building giant warehouses full of AI chips on Earth and instead put those chips on satellites, powered directly by the sun and connected to each other by laser beams. Reports on May 12 and 13, 2026, said Google is now in talks with SpaceX to actually launch the hardware. A small test mission with two satellites is planned for early 2027. The longer-term goal is a tight formation of 81 satellites circling within roughly one kilometre of each other, behaving like a single data centre in orbit. The catch is cost: today, sending a kilogram of equipment to space costs more than ten times what the maths requires. The whole plan only works if launches become much cheaper and much more frequent.

On a Mountain View campus where a senior director named Travis Beals first pitched Suncatcher to colleagues as a “moonshot” in late 2025, the slide deck has quietly turned into a procurement conversation. Last week, the Wall Street Journal reported that Alphabet is in commercial talks with SpaceX over launch services for Project Suncatcher, the orbital compute programme Beals and Blaise Agüera y Arcas first detailed in a public Google Research preprint titled “Towards a future space-based, highly scalable AI infrastructure system design.” Bloomberg and TechCrunch confirmed the talks within hours. Google had previously named Planet Labs as the satellite bus partner and acknowledged it was speaking with multiple launch providers; SpaceX is now, on the record, one of them. The choreography matters. Alphabet already owns roughly 6.1% of SpaceX as of year-end 2025, making this less an arm’s-length tender than a corporate handshake between two entities that share a balance sheet exposure to the same outcome. The proposed architecture is concrete: a constellation of solar-powered satellites in a dawn–dusk sun-synchronous low-Earth orbit at roughly 650 km, each carrying Google’s Trillium-class v6e TPUs, stitched together by free-space optical interconnect at up to 1.6 terabits per second per transceiver pair. The 81-satellite cluster is engineered to fly inside a one-kilometre radius using formation control modelled on the Hill–Clohessy–Wiltshire equations. Beals’s team has the receipts on the physics. Trillium TPUs were placed in a 67 MeV proton beam at the Indiana University cyclotron and survived a cumulative dose of 2 krad(Si) before the High-Bandwidth Memory began to misbehave — almost three times the shielded five-year dose Google projects for the mission. “The High Bandwidth Memory subsystems were the most sensitive component,” the paper notes, “but no hard failures were observed.” The radiation argument, the one that has killed every previous data-centre-in-space pitch since the 1990s, is no longer the central blocker. The blocker is launch economics. Google’s internal model puts the break-even point for orbital compute at $200 per kilogram delivered to low-Earth orbit. Current Falcon 9 commercial pricing runs $1,500–$2,900 per kilogram. Closing that gap requires SpaceX’s Starship to fly at roughly 180 missions per year by the mid-2030s — a cadence that exists today only in Elon Musk’s slide decks and in Gwynne Shotwell’s manifest. Google’s own published target year for $200/kg is 2035. Until then, by the company’s own admission in adjacent analysis, orbital infrastructure is roughly three times more expensive than its terrestrial equivalent. The news landed in a market already primed for it. On May 12, the same day the WSJ report broke, the Clark Capital 60% RSI capex signal flagged hyperscaler infrastructure intensity at levels last seen in 1999. The two stories are not coincidence; they are the same story, told from opposite ends of the cost curve.

Start with the capex denominator. Microsoft, Alphabet, Amazon and Meta have collectively guided to roughly $725 billion in 2026 capital spending, up about 77% year-on-year, with roughly three-quarters earmarked for AI-specific infrastructure. That single-year figure is larger than the entire 2025 GDP of Belgium, and it is being deployed against grid interconnect queues that, in the PJM and ERCOT zones, are now quoted in five-to-seven-year windows. Microsoft alone disclosed an $80 billion backlog of Azure orders it cannot fulfil because it cannot find the megawatts. The economic case for Suncatcher is therefore not that space is cool. It is that the terrestrial supply curve for power has gone vertical. The Suncatcher unit economics are a wager on two compounding curves. The first is launch cost. SpaceX has promised that Starship, fully reusable, will undercut Falcon 9 by up to 90%. Falcon 9 itself already cut launch prices roughly 95% versus the Space Shuttle. If Starship delivers even half of its stated improvement and flies at the cadence Musk has guided, $200/kg by 2035 is aggressive but not absurd. The second is power. A satellite in dawn–dusk sun-synchronous orbit collects sunlight for roughly 99% of every revolution, against a terrestrial solar capacity factor closer to 22%. Per square metre of panel, the orbital harvest is therefore around four-and-a-half times the ground equivalent — before transmission losses, intermittency penalties, and the standing fight over grid hookups in Frankfurt and Northern Virginia. The constellation maths is equally specific. Eighty-one satellites at one-kilometre separation, connected by 1.6 Tbps optical links, yields aggregate intra-cluster bandwidth in the petabit range — roughly comparable to the spine of a single hyperscale availability zone today. Google is not, in 2027, planning to retire Council Bluffs. It is planning to prove that one rack-equivalent of TPUs can be flown, powered, cooled by radiator surface area, and addressed by a scheduler that treats a 600-kilometre-altitude bus as just another node in Borg. More remarkable still is the supply chain it implies. Starcloud, an Nvidia-backed Redmond startup, already put an H100 in orbit in November 2025 aboard Starcloud-1 and trained a model on it in December — a $1.1 billion valuation predicated on the same thesis. SpaceX has separately filed with the FCC to fly up to one million of its own data-centre satellites. The orbital compute layer is not a single bet; it is becoming a market structure.

The architectural choices Google has made tell DAX40 CIOs more about the next decade of cloud than the headline does. Suncatcher is not a small number of large satellites; it is a large number of small ones, flying close enough together that the optical interconnect behaves like a backplane rather than a WAN. That is a deliberate copy of the design pattern that made TPU pods dominant on Earth: tight coupling, low latency, all-reduce at the speed of light. It is also the design pattern that lets Google substitute optical photons for the copper interconnect that currently constrains rack density in Council Bluffs and Eemshaven. The thermal story is the unresolved one. In vacuum there is no convection; every watt of TPU heat must leave the spacecraft as infrared radiation through a deployable radiator. Google’s paper sidesteps the per-satellite power envelope, but third-party physicists have noted that radiator mass scales roughly linearly with dissipated power, which puts pressure on the very launch-mass budget that drives the $200/kg threshold. The cooling problem does not kill Suncatcher; it caps the cluster’s effective power density well below a terrestrial GB200 rack. Suncatcher is not a like-for-like replacement for a hyperscale region. It is a different shape of compute. The other open question is debris. Eighty-one satellites flying inside a one-kilometre radius is a formation; a million satellites — SpaceX’s FCC ceiling — is a Kessler-syndrome conversation. Google published a separate debris-risk analysis in January 2026; the European Space Agency and the German Bundesnetzagentur have not yet been heard from in detail. The regulatory perimeter around Suncatcher is, today, a blank page.

For two decades, enterprise software earned its keep by being the place where humans clicked. Salesforce became the screen for sales reps, Workday the screen for HR, SAP the screen for a controller closing the books. The screen was the lock-in: train ten thousand users, integrate twenty workflows, and switching costs did the rest. Agentic AI breaks that pattern. When a software robot reads and writes directly to the underlying database through an API, the screen stops mattering. A May 2026 essay from venture firm Andreessen Horowitz argues this strips classic SaaS of its defences and forces the moat to move — down into data, permissions and compliance, or up into networks and real-world execution. For any CIO weighing a multi-year SAP, Workday or Salesforce commitment, the question is suddenly load-bearing.

On a Wednesday morning in mid-May, a partner at Andreessen Horowitz hit publish on an essay with a deliberately provocative title — “Is Software Losing Its Head?” — and within hours it was circulating through CIO Slack channels from Munich to Walldorf. Seema Amble, who has spent the last two years writing a16z’s most-read pieces on enterprise software, framed the argument with surgical clarity: “In the SaaS era, the system of record was defensible because humans lived in the interface. In the agentic era, that advantage weakens.” The defensible layers, she wrote, shift downward into data models, permissions, workflow logic and compliance, and upward into networks, proprietary data generation and real-world execution. Everything in the middle — the UI, the muscle memory, the workflow ergonomics that two decades of SaaS pricing was built on — is now contested ground. The essay did not arrive in a vacuum. Four weeks earlier, at TrailblazerDX 2026 in San Francisco, Marc Benioff had walked on stage and unveiled Salesforce Headless 360. “Our API is the UI,” he posted afterwards. “Entire Salesforce, Agentforce and Slack platforms are now exposed as APIs, MCP and CLI.” More than sixty new Model Context Protocol tools shipped that week, alongside an Experience Layer that decouples what an agent does from how it appears — the same rebooking card can now render inside Slack, ChatGPT, Claude or Microsoft Teams without Salesforce’s browser ever loading. For the company that taught a generation of enterprises to live inside a tab labelled lightning.force.com, this was a doctrinal reversal. The pivot is the part that should worry incumbent vendors. Headless 360 is being marketed as the inevitable next step for the platform leader, but it is also a confession: if the agent is the user, the UI cannot be the lock-in. Amble’s essay reads the move as a hedge — Salesforce betting that its value sits in the data layer rather than the screen, and trying to be the one to commoditise its own front-end before someone else does it for them. Across the Atlantic, SAP arrived at Sapphire 2026 in Orlando with the opposite gesture. Christian Klein opened his keynote with a line aimed squarely at the same anxiety: “No AI agent can compensate for a bad data landscape.” Behind him, slides announced more than forty specialised Joule agents, 2,400 Joule Skills, a Business AI Platform consolidating BTP, Business Data Cloud and Business AI into a single stack, and — most consequentially — a revised API policy that explicitly prohibits using SAP APIs for interaction or integration with semi-autonomous or generative AI systems that plan, select, or execute sequences of API calls unless the agent is SAP’s own. Two of the three largest enterprise vendors in the world, in the same six-week window, executed mirror-image strategies built on the same premise: in the agentic era, only the data moat survives.

Strip away the marketing, and the architecture question reduces to a single diagram. Underneath every system of record sits a schema — customers, opportunities, employees, journal entries, purchase orders — layered with permissions, validation rules, audit trails and compliance logic. Above the schema sits an application server enforcing workflow. Above that, a UI. For three decades, vendors priced and defended the package. Agents detonate the package by hitting the schema directly via API or MCP, leaving the UI and much of the workflow runtime as optional. The architectural choice for a CIO becomes: which of those layers do you still want to buy from a single vendor, and which do you stitch together? a16z sketches three paths. Path one: ride the incumbent. Keep SAP or Salesforce as the system of record and consume agents through the vendor’s endorsed product — Joule, Agentforce, Workday Sana. Path two: build your own system of record, with custom schemas, permissions and agents constructed from scratch on cloud primitives. Path three: replace the incumbent with an AI-native challenger designed from the first commit for machine readability and agent orchestration. Each path implies a different bet on where defensibility actually lives. The historical comparison is the early-1970s mainframe wars. Back then, IBM’s lock-in was the operating system and the leased hardware; when the PC and TCP/IP unbundled those, the moat moved into databases (Oracle), middleware (BEA, Tibco) and eventually applications (SAP, Siebel). Each unbundling event wiped out roughly a generation of incumbents and produced a generation of challengers. The SaaSpocalypse of January 2026 — in which one trading session erased roughly 300 billion dollars in software market value after Anthropic’s Claude Cowork launch — has the same shape: investors marking down the part of the stack that is about to be exposed. Three concrete shifts follow for DACH CIOs. First, contract leverage tilts towards the buyer for the first time in a decade; Fortune’s April 2026 coverage of the harder line CIOs are taking with vendors is already showing up in DAX40 renewal cycles. Second, the data layer becomes a procurement criterion in its own right — SAP’s restrictive API policy is rational from a vendor view but creates a genuine compliance problem for any enterprise running a multi-vendor agent strategy against S/4HANA data, as analyst Kai Waehner has documented. Third, the experience layer fragments: Headless 360’s decoupling means the same Salesforce record can be acted on inside Microsoft Teams or Claude, which is liberating for users and existential for vendors whose pricing was tied to seats in their own UI. The architectural decisions made in the next twelve months will determine which side of the moat a company sits on for the next ten.

Since October 2022, Washington has tightened the screws on which Nvidia chips can legally cross into China. The H100, the A100 and their export-tuned cousins are off-limits; the watered-down H20 comes and goes with each diplomatic season. The intent was straightforward: starve Chinese labs of the silicon needed to train frontier models. Three and a half years on, the picture is more complicated. China’s compute stock still trails the United States by perhaps two to three years, but its open-weight models, Qwen, DeepSeek, Kimi, GLM, Doubao, run only six to eight months behind the closed US frontier and at a fraction of the inference cost. Azeem Azhar’s Exponential View essay, published on the morning Donald Trump’s plane touched down in Beijing, argues that scarcity, not abundance, has become China’s competitive teacher.

In a windowless room in Hangzhou, an engineer slides a laptop across the table and shows Azeem Azhar a profiling dashboard. Every kernel is annotated, every memory transfer accounted for to the millisecond. “We do not have the luxury of waste,” the engineer tells him, gesturing at a rack of H800s, a generation behind what his Californian counterparts take for granted. “So we treat each GPU-hour as if it were a barrel of oil.” It is a phrase Azhar will hear, in different words, in eleven of the fourteen labs he visits over seven days: DeepSeek and MoonshotAI in Hangzhou and Beijing, MiniMax and Z.ai in Shanghai, ByteDance, 01.AI, Alibaba, Ant Group, Xiaomi, AInnovation, Galbot, Unitree, ModelScope and RWKV. The result, published on May 13 as Inside China’s AI labs: the efficiency moat, lands precisely as Air Force One descends on the Chinese capital with Jensen Huang, Tim Cook and Elon Musk among the entourage. The timing is not subtle. Azhar’s central claim is that US export controls have inadvertently turned Chinese labs into what he calls ruthlessly efficient engineering shops, extracting four to seven times more useful intelligence per unit of compute than naive scaling laws would predict. The mechanism is part necessity, part culture. Forbidden the brute-force option, Chinese teams have rewritten attention kernels, embraced aggressive mixture-of-experts sparsity, leaned into FP8 training, fused communication and computation, and shared optimisations promiscuously across an open-weights ecosystem that now spans more than 100,000 derivative models on Hugging Face. The narrative pivot is sharp. For two years the working assumption inside DAX40 boardrooms and at Anthropic-adjacent policy salons in Washington was that compute would compound, that the Bay Area’s GPU stockpile would translate, almost arithmetically, into model superiority. The opposite has happened on the open-weights tier. Qwen’s family is now the most-downloaded open model lineage on the planet. DeepSeek V4, released in late April, runs on Huawei Ascend 950 supernodes and Cambricon silicon rather than smuggled Blackwells. Kimi K2.5 from MoonshotAI posts 76.8 percent on SWE-bench and 74.9 percent on BrowseComp, numbers that were exclusive to closed US labs a year ago. Annual recurring revenue at Moonshot crossed two hundred million dollars in April, and Airbnb, an American company, has quietly moved its customer-service chatbot stack to Qwen. The historical parallel Azhar reaches for is the 1970s Soviet aviation industry, which, denied Western avionics, learned to build airframes that were crude but uncannily fuel-efficient. ‘Constraint is the mother of architecture,’ he writes, paraphrasing a DeepSeek researcher who asked not to be named. The point is not that China has caught the United States. It has not. The point is that the export-control regime, designed to widen the gap, may have narrowed the gap that actually matters to enterprise buyers.

Strip away the politics and the arithmetic is what should preoccupy a DACH chief technology officer. On a per-token basis, Chinese API endpoints are pricing five to thirty times below the US frontier as of April 2026, according to comparative pricing data compiled by Bloomberg and corroborated in independent benchmark scrapes. DeepSeek V4 sits at roughly one quarter of the inference cost of GPT-5.4 for benchmark-comparable workloads, a figure Azhar cites and which broadly tracks numbers Nathan Lambert has been publishing at Interconnects through the year. Compute stock tells the other half of the story. The United States hyperscaler triumvirate, Microsoft, Google and Amazon, will deploy on the order of five to six million high-end GPU equivalents by year-end 2026, against a Chinese installed base estimated at roughly one to one and a half million, much of it H800-era silicon or domestic Ascend 910C and 950 parts. That is the two-to-three-year compute gap Azhar refers to. But the conversion ratio, intelligence delivered per FLOP consumed, runs the other way. His four-to-seven-times figure is back-of-the-envelope, derived from comparing reported training FLOPs against MMLU, GPQA, SWE-bench and a basket of agentic evaluations, and he is careful to flag the uncertainty. Even at the lower bound, the implication is uncomfortable for the prevailing US capex thesis. If a Chinese lab can match a US lab’s open-tier capability with one-fifth the compute, then the marginal dollar of hyperscaler capex purchases less competitive moat than the bull case assumes. Anthropic’s Dario Amodei has pushed back on exactly this framing throughout 2026, arguing in repeated congressional testimony and on his personal blog that DeepSeek and its peers follow the expected cost-reduction curve rather than break it, and that the frontier, the genuinely novel capability, still requires the millions of chips and tens of billions of dollars that only the US ecosystem can muster. The counter-argument, voiced in Brussels and increasingly in Berlin, is that the frontier is not what most enterprise workloads need. A risk-classification pipeline at a German Landesbank does not require GPT-5.4-level reasoning; it requires reliable, auditable, cheap inference on European infrastructure. On that test, the Chinese open-weight stack now wins on three of four axes and loses, decisively, on the fourth: jurisdictional trust. Italian, French, Belgian and Irish regulators have all opened investigations into DeepSeek’s data handling under GDPR; the German BfDI has signalled it is watching. The pricing arbitrage, in other words, exists. So does the regulatory friction. The interesting question for 2026 is which one compounds faster.

Until this week, the corporate laptop market had two AI stories. Microsoft was retrofitting Windows with Copilot and badging compliant hardware as Copilot+ PCs. Apple was layering Apple Intelligence on top of macOS for its premium minority. Everyone else — the ChromeOS schools, the Linux holdouts — was background noise to a CIO buying tens of thousands of endpoints from Dell, HP or Lenovo. On 12 May 2026, at the Android Show pre-I/O event, Google announced Googlebook, a new category of Gemini-native laptops co-engineered with those same five enterprise OEMs. The pitch: an operating system rebuilt as an intelligence system, with an AI-aware cursor, prompt-built widgets and tight Android-phone handoff. Devices ship in autumn. The strategic point is not the gadget. It is that the firm that ships Search and Gemini now has hardware partners aimed at the very fleet CIOs renew every four years.

A short video opened the stream: a slate-grey laptop, lid closed, with a thin strip of light along the hinge — the Glowbar — pulsing softly as if breathing. Cut to Alex Kuscher, Google’s senior director for laptops and tablets, on a sound-stage in Mountain View. “Over 15 years ago, we introduced the Chromebook, a laptop built for a cloud-first world,” he said. “Now, as we are moving from an operating system to an intelligence system, we see an opportunity to rethink laptops again.” With that sentence, Kuscher pensioned off the Chromebook brand in everything but name and unveiled Googlebook — a category, not a single SKU — built on a fusion of Android and ChromeOS underpinnings, with Gemini wired into the input layer. The demos were aimed squarely at knowledge work. Wiggle the cursor and a contextual halo opens around it: this is Magic Pointer, built with Google DeepMind. Hover a date in Gmail and it offers to draft a calendar invite; pair two images on a moodboard and it composites them; highlight a paragraph and it rewrites, translates or summarises. Create your Widget takes a natural-language prompt and assembles a live dashboard from Calendar, Gmail, Maps, Drive and the open web — the prompt-to-UI loop Microsoft has been demoing for two years but rarely shipping. Cast my apps lets a user run an Android phone app, in a window, on the laptop, using the phone’s install and entitlements. A Quick Access file browser exposes phone storage to the desktop without sync. The pivot came when Kuscher rattled off five names. Acer, Asus, Dell, HP, Lenovo — the same five vendors that fill every IT procurement RFP from Frankfurt to Singapore — will all ship Googlebooks this autumn. Every device must carry the Glowbar, an opinionated piece of industrial design that doubles as a status light for Gemini activity. Google did not show pricing, silicon partners or enterprise management tooling, and the keynote was studiously silent on Workspace integration. But the strategic message landed before the questions could be asked: the company that owns the world’s most popular browser, the world’s most popular mobile OS and one of the two leading frontier-model families has now licensed those advantages to the major corporate PC channel. That is what makes this announcement different from every previous attempt at a non-Windows enterprise client. Earlier ChromeOS pushes — the 2017 Chromebook for Enterprise programme, the 2020 Pixelbook Go — leaned on Google’s own hardware and education-channel goodwill. Googlebook leans on Dell, HP and Lenovo, whose account managers already sit inside DAX40 procurement organisations and whose product roadmaps are now, by their own choice, bifurcated between Windows+Copilot and Android+Gemini. For an industry whose default answer to which laptop to buy has been the next Lenovo ThinkPad refresh for two decades, that is the news.

The historical comparison that frames Googlebook is not Chromebook. It is Microsoft’s own Copilot+ PC launch at Build 2024. Then, as now, an incumbent platform owner co-opted the major OEMs to brand a new category of AI-first machines, with custom silicon requirements, premium positioning and an OS-level AI feature (Recall) as the differentiator. The launch did not, in Engadget’s blunt phrase, catch fire. Independent tracker data through Q2 FY26 shows Microsoft Copilot for M365 at roughly 15 million paid seats — fast sequential growth, but only about 3.3% of the addressable installed base, and Recon Analytics has the product’s accuracy NPS in negative territory. When enterprises offered employees a choice of assistants, ChatGPT won 76% of voluntary usage, Copilot 18% and Gemini 6%. Microsoft has since quietly pulled Copilot out of several Windows surfaces after user backlash. That is the run-rate Google is trying to attack. The second context point is the OS unification story. Google has spent the last eighteen months publicly signalling that Android and ChromeOS would converge — a project Hiroshi Lockheimer’s successors confirmed last summer. Googlebook is the commercial expression of that engineering merge: a single OS image that can run Play-store apps natively, host a Chromium browser as a first-class shell and surface Gemini at the input layer rather than as an app. The technical implication is more important than the marketing one. Android already runs on roughly three billion active devices and ships with a mature MDM surface (Android Enterprise) that Samsung Knox, Microsoft Intune and VMware Workspace ONE all integrate against. A Googlebook fleet, in theory, drops into the same management plane a CIO already uses for corporate phones — a meaningful operational concession compared with running parallel Windows, macOS and mobile estates. The third context point is competitive timing. Apple’s WWDC keynote is on 8 June. Apple Intelligence, launched eighteen months ago, has under-delivered relative to the keynote; the new Siri remains delayed and on-device model quality has lagged Gemini Nano and Phi-4. With Googlebook now public, Apple faces an awkward keynote: any narrowing of the Apple Intelligence story will read as a retreat, while any ambitious expansion will be measured against demos the rival has already shipped on stage. Microsoft’s Build conference, which begins next week in Seattle, has the same problem inverted — the Copilot+ PC narrative now competes with a vendor that controls the browser, the phone and the model. For DAX40 CIOs the procurement consequence is concrete. Every device standard written before this week assumed two endpoint AI stacks worth tracking. From autumn there are three, and the third arrives bundled with Dell, HP and Lenovo — vendors who will offer Googlebooks alongside Windows on the same quote. Decisions that were postponable in 2025 — which AI assistant becomes the corporate default, which model family gets the Workspace or Microsoft 365 contract, which endpoint MDM survives consolidation — will need to be made in the 2027 budget cycle. The Android Show was a fifty-minute keynote. The procurement-cycle ripple will last years.