The Redline Report: Frenemy may be the word of the decade. The age of owning the whole AI stack is over.
Speed Read
- Every week brings an AI deal that would have looked like treason five years ago. One at a time they are trees. Together they are a forest: the giants of AI now run on each other.
- The cause is scale. Anthropic’s annualized revenue run-rate reportedly rose about fivefold in the first five months of 2026. No one company can build the chips, power, models, agentic applications and distribution fast enough alone.
- Six deals show the pattern: SpaceX and Nvidia, SpaceX and Anthropic, Microsoft and Anthropic, Amazon and OpenAI, Apple and Google, and Meta with nearly everyone.
- Owning the whole stack is finished. What stays proprietary is your map of your own business. Everything else is rented, routed or shared.
- Enterprises must follow: multi-model by default, partners across the stack and, where a problem is bigger than any one firm, alliances with their own competitors.
The assignment: make a friend of a rival before your rival does.
The frenemy decade
On Wednesday, October 7, Elon Musk posted an “important note” that would have been unthinkable a year ago. Grok Bot, the agent his company builds with Cursor, would “use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno.” Translation: the man who owns a frontier AI lab told his customers that sometimes the best answer comes from a competitor’s brain.
That is a bit like Chevy telling its dealers to send a customer to Ford when the customer needs a pickup. And it was not even the strangest thing in Musk’s week. The day before, Reuters reported that SpaceX was lining up about $40 billion of debt to buy more Nvidia chips. Some of the chips it already owns sit in Memphis data centers rented to Anthropic, reportedly for $1.25 billion a month. SpaceX is now Anthropic’s landlord, its customer and its competitor, all at once.
Nobody blinked. That is the story.
Look at any single AI headline and you see a tree: a chip order, a cloud contract, a model license. Step back and you see the forest. The largest companies in technology have stopped trying to own the whole stack. They buy from, sell to, invest in and route work through the very companies they compete with. “Frenemy” is slang for a friend who is also an enemy. In AI it has become the operating model, and it could end up being the word of the decade.
This piece walks through six recent deals that show it most clearly, explains why it is happening now, and makes the case that every enterprise has to follow: across models, across partners, and in some cases across the table from its oldest competitors.
Part I: The trees
Six deals, each surprising on its own. Read them in a row and the surprise wears off, which is exactly the point.
1. SpaceX × Nvidia: Minihards on the ground, Starminds in the sky
Musk names his Memphis data centers the way a teenager names a garage band. The first two are Macrohard and Macrohardrr, an unsubtle poke at Microsoft. The newest is Minihard.
Don’t let the name fool you. Minihard packs the same 220,000 Nvidia GB300 GPUs as Macrohardrr into roughly a quarter of the floor space. It is built as a standard module, identical in layout, so software tuned for one unit runs unchanged on the next. Colossus 1 reportedly went from empty building to operational in about 122 days, against the 18 to 24 months a conventional data center takes. SpaceX is turning AI factories into a production line, the same way it turned rockets into one. Henry Ford would be proud.
The second front is in orbit. On August 4, SpaceX announced it is co-designing its Starmind AI1 satellite with Nvidia, each one carrying Rubin GPUs and Vera CPUs for “datacenter class space compute.” Musk expects launches to start in late 2027 and reach significant scale in 2028, with an ambition of up to a million orbital data centers.
The money runs both ways. Nvidia took equity in xAI’s $20 billion round in January; those shares became SpaceX stock when the two companies merged. On October 6, Reuters reported that SpaceX was raising about $40 billion of debt, led by Apollo, to buy more Nvidia chips, and Musk said SpaceX will use Nvidia hardware exclusively in its data centers.
- What Nvidia gets: a customer that builds its own buildings, power and launch vehicles at production-line speed, plus a shareholder’s seat in the room.
- What SpaceX gets: the one supplier able to fill a Memphis campus and a sky full of satellites.
Point of view: This is vertical integration by partnership. Neither company reaches this scale alone, and each is betting that the other’s growth is its own. The old playbook said own the factory. The new one says co-design it with your supplier, and let the supplier own a piece of you.
2. SpaceX × Anthropic: landlord, customer and competitor
In May, Anthropic agreed to use the entire Colossus 1 data center, more than 220,000 Nvidia GPUs and over 300 megawatts, to keep up with demand for Claude. Analysis of SpaceX’s IPO filing puts the deal, which extends to Colossus II, at about $1.25 billion a month through May 2029. Either side can walk away on 90 days’ notice. Anthropic has also said it wants to work with SpaceX on “multiple gigawatts” of orbital compute.
Then came October 7. Grok Bot, the agent SpaceX launched in beta in August with Cursor (which SpaceX bought for $60 billion in stock), would route each task to whichever model does it best, Claude Opus 5.5 included. Musk’s stated goal: “Whatever is most likely to give you the best outcome.”
So the question of Opus or pure Grok has been answered by the man with the most reason to say pure Grok. The answer is both, task by task. Said more simply in Spanish, “Mi casa es su casa, y su casa es mi casa.”
The arrangement is wonderfully circular. SpaceX rents its GPUs to Anthropic. Anthropic serves Claude on them. SpaceX’s own agent then pays to call Claude. Nvidia owns a slice of SpaceX and sells chips to everyone in the loop. Somewhere an antitrust lawyer is drawing this on a whiteboard and running out of arrows.
Point of view: Loyalty to a single model is now a luxury nobody can afford, not even the people who make models. The winning agent is the one that knows which brain to call for which job. That is a routing decision, and routing is only as good as the agent’s understanding of the work in front of it.
3. Microsoft × Anthropic: the house that OpenAI built
Microsoft spent the better part of a decade, and more than $13 billion, making OpenAI the engine of its AI strategy. In March it launched Copilot Cowork, its most ambitious agent yet, built with Anthropic’s Claude. It sells inside a new $99-per-user Microsoft 365 E7 tier, 65% above the E5 price. For bonus irony, Anthropic already sells its own product called Cowork.
The relationship runs deeper than one product. Last November, Anthropic committed to buy $30 billion of Azure compute, while Microsoft and Nvidia agreed to invest in Anthropic. Microsoft’s stated policy is now to pick the best model for each task, whoever built it. Musk named two buildings after mocking Microsoft. Microsoft, it turns out, was busy doing the same thing he was: shopping for the best brain.
Point of view: The most powerful software distributor on earth decided its customers’ results matter more than its loyalty to one model. Microsoft kept what it is uniquely good at, the distribution, the identity layer and the context sitting in your inbox and calendar, and rented the intelligence. Own the context, rent the brain. That is the template.
4. Amazon × OpenAI: the double date
Amazon is Anthropic’s largest backer, and Claude is the anchor model on AWS. In February, Amazon also committed $50 billion to OpenAI: $15 billion up front and $35 billion more once conditions are met.
In return, OpenAI agreed to use about 2 gigawatts of Amazon’s in-house Trainium chips, expanding an existing $38 billion AWS agreement by $100 billion over eight years. AWS became the exclusive third-party cloud distributor of OpenAI Frontier, OpenAI’s platform for running teams of agents. The two are co-building a runtime for agents on Amazon Bedrock, which now sells OpenAI and Anthropic models side by side. Andy Jassy sounded especially pleased that OpenAI chose to “go big” on Amazon’s own silicon.
Amazon used to be in a committed relationship. Now it has a portfolio.
Point of view: Amazon is hedging, and the hedge is the strategy. In a market that reshuffles every six months, the safest place to stand is wherever all the models meet. For AWS that place is Bedrock. For an enterprise it is an architecture that can swap one model for another without rewriting the business.
5. Apple × Google: Siri gets a new brain
Apple builds its own chips, its own operating systems and, famously, its own everything. In January, Apple confirmed that the next-generation Siri and the Apple Foundation Models behind it would be built on Google’s Gemini. Google now describes itself as Apple’s “preferred cloud provider,” and Apple has reportedly asked Google to look at hosting Siri servers in Google’s own data centers.
This is the same Google that pays Apple billions a year to be Safari’s default search engine, and the same Google whose Android phones Apple has spent two decades trying to beat. Now the money flows in both directions, and Apple’s most personal product thinks with its biggest rival’s brain.
Point of view: If Apple, the patron saint of vertical integration, will rent the intelligence behind Siri, no company can still claim that owning every layer is a strategy. Apple kept the device, the privacy promise and the customer relationship. It rented the one part it could not build fast or well enough. Knowing which part is which is the whole game.
6. Meta × everyone: renting from rivals, opening up Muse
Meta is the most interesting case, because it is collaborating in every direction at once.
- Chips from a rival. In February, Meta signed a multi-year, multi-billion-dollar deal to rent Google’s TPU chips, days after announcing giant GPU orders from Nvidia and AMD. Google competes with Meta for every advertising dollar on earth. It is now also a supplier.
- Open Muse, built with Apple. On October 15, Meta released Muse Spark Open as an open-weights model anyone can download. Mark Zuckerberg thanked John Ternus, Apple’s CEO since Tim Cook handed over the job on September 1, by name for months of joint work tuning it to run on linked Mac Studios. It is the first public partnership between the two companies in close to a decade, after years of open warfare over privacy and ad tracking.
- An open door for partners. Muse, Meta’s personal agent, launched in September and already checks out at Walmart, Best Buy, Sephora, Gap, Wayfair and Expedia, with payments through Stripe, Shopify and PayPal. Meta opened the platform so outside developers can build their own connectors and received more than 1,500 applications in under a week.
The business model explains the generosity. Muse is free for a large number of tokens, and Meta expects to earn a small fee on transactions over time. Every partner that plugs in makes the agent more useful, and every rival that helps distribute it widens the funnel. Not everyone has signed up: Amazon blocked Muse from shopping its store. Even frenemies have a guest list.
Point of view: Meta has learned that openness is a distribution strategy, not a charity. Release the model, open the connectors, partner with the company you fought for a decade, and let the ecosystem do the selling. The walled garden still has walls. It just has a lot more gates.
Part II: The forest
Put the six deals on one page and the same handful of names keeps appearing as each other’s supplier, customer, investor, landlord and competitor. This is not a series of exceptions. It is the new structure of the industry, and it is not the cloud or mobile era replayed. Those platforms rewarded whoever owned the ecosystem: the app store, the operating system, the cloud region. AI rewards whoever can assemble the most capability fastest, and no one can do that alone, because every critical input (chips, power, models, agentic applications, data, distribution) is scarce and held by someone else. Three forces hold the new structure in place.
The scale is beyond any one balance sheet. Figures reported around Anthropic’s confidential IPO filing put its annualized revenue run-rate at about $47 billion in late May 2026, up from roughly $9 billion at the end of 2025. Those numbers count cloud-partner sales gross and are unaudited, but even heavily discounted, the curve is steeper than anything in enterprise software history. Morgan Stanley estimates that AI infrastructure will need $1.5 trillion of outside financing by 2028. SpaceX has gone public to help fund it, and Anthropic has filed to follow. When demand grows that fast, the constraint is capacity, and capacity is wherever you can find it, including at your competitor.
The plumbing is becoming shared. Last December, Anthropic, OpenAI and Block placed their agent standards, MCP, AGENTS.md and goose, into the Linux Foundation’s new Agentic AI Foundation. The rivals agreed on the pipes so they could compete on what flows through them. Once agents speak a common language, any agent can call any model, any tool and any partner.
Partnerships are built to be unbuilt. The Colossus agreement can end on 90 days’ notice. Microsoft is no longer exclusive to OpenAI, and OpenAI is no longer exclusive to Azure. These are deep partnerships on short leashes. That is the working definition of frenemy: cooperate fully today, keep the freedom to compete tomorrow.
The result is that proprietary ownership of the whole stack is over. What remains proprietary is narrower and far more valuable: the data, context and know-how that describe a company’s own business. Models supply the reasoning. The business supplies the grounding. Look at what each giant kept. Microsoft kept the inbox and the calendar. Apple kept the device and the customer. Meta kept the social graph. SpaceX kept launch, power and speed of construction. Everything else, they are willing to rent, share or route to someone else.
Part III: The Coke and Pepsi test
The giants have shown what frenemy looks like at the top of the AI stack. The more important question is whether the same logic holds one layer down, in the industries that buy AI rather than build it. It does, and the first movers are already there.
We will probably never see Coke selling Pepsi. But AI is already pushing direct competitors to cooperate on the problems none of them can solve alone. The rule of thumb: compete on what makes you different, collaborate on what doesn’t.
Pharma: rivals training one model
Five drugmakers, AbbVie, Astex, Bristol Myers Squibb, Johnson & Johnson and Takeda, jointly trained the OpenFold3 protein model on more than 20,000 of their own confidential structures. Not one raw structure changed hands; only model parameters moved, through a federated system run by Apheris. High-quality protein-ligand predictions rose from 35.6% to 52.1%, and training took under ten weeks. A follow-on project with AstraZeneca launched in September. As Apheris’s CEO put it: “In drug discovery, no single company has enough data to solve the hardest problems alone.”
Hospitals and payers: from bot wars to a shared rulebook
Healthcare claims have become an arms race. Payers have used automation to scrub and deny claims for years, and hospitals have answered with AI of their own. In September, the Blue Cross Blue Shield Association reported that hospitals were systematically adding secondary diagnoses, raising payments by nearly $12,000 per case on average. Hospitals say payer algorithms drive denials. The result is one side’s bots arguing with the other side’s bots, and both sides paying for the argument.
The frenemy alternative is already on the calendar. Federal rules require payers to offer electronic prior-authorization interfaces from 2027, and more than 50 insurers pledged in 2025 to make most prior-authorization decisions in real time. The next step is a shared, machine-readable rulebook, the codes, coverage policies and clinical evidence, that both sides’ agents read the same way. Competing hospital systems could go further and pool their denial data, as the drugmakers did, to train models no single system could build alone.
Broadcasters and publishers: rivals writing the rules together
In February, the BBC, the Financial Times, the Guardian, Sky News and the Telegraph, rivals for every reader and every scoop, formed SPUR, a coalition to set common standards for licensing their content to AI companies, and invited broadcasters and publishers worldwide to join. In the US, publishers including Reddit, Yahoo, Medium and People Inc. backed RSL, a shared standard for telling AI crawlers what content costs. Alone, a newsroom is a website waiting to be scraped. Together, publishers are a supply market with terms.
Carmakers: pooling the plumbing
Volkswagen, BMW, Stellantis and Mercedes-Benz, with Bosch, ZF and many other suppliers, formed an open-source software alliance to pool the code that does not set their cars apart. They compete on design, performance and brand, not on the operating system under the dashboard.
Banks: subsidizing the fraudsters
Banking is the cautionary tale. Each large bank spends heavily to build its own fraud models, each trained on its own small slice of the crime. The criminals have no such handicap: fraud rings trade stolen identities, scripts and attack methods in open forums, in real time. The US Treasury has said banks share too little fraud data for AI to work well, and warned that smaller institutions fall furthest behind.
That makes the industry’s caution expensive. Refusing to federate is not risk management. It is an operating subsidy to the fraudsters. The method the drugmakers used, where the model travels and the raw data never leaves the building, is exactly what a fraud consortium needs. The bank that organizes it first will set the terms for everyone else.
Part IV: The frenemy playbook
What the giants did at planetary scale, an enterprise can do at its own. Five moves:
- Go multi-model by default. If Musk won’t marry Grok, you shouldn’t marry anyone. Route each task to the model that does it best, and expect the best model to change twice a year.
- Own the map, rent the territory. The map is your semantic or ontology layer: the codified rules, the relational workflows and the context graph that describe how your business actually creates value, from who your customers are to what your contracts allow and how work moves from one desk to the next. Models provide reasoning; the map provides grounding. Rent the reasoning freely. Rent the grounding too, and you have outsourced your moat. With that map, any model can do useful work. Without it, even the best model is a tourist.
- Find your federation. List the problems your whole industry shares and nobody wins alone: fraud, safety, compliance, claims, training data. Those are the problems to solve with your competitors.
- Govern the borders. Openness is only safe with fences: agents that stay in their sandboxes, every action traced, every cost counted, every decision explainable. Good fences make good frenemies.
- Write the exit into the entry. Build partnerships the way the Colossus deal was built: deep enough to matter, short enough to leave. A 90-day termination clause is not distrust. It is the price of moving fast.
The executive matrix
The whole argument reduces to three decisions, made layer by layer.
| Strategy | Layer | Enterprise action |
|---|---|---|
| Rent and route | Foundation models, compute, infrastructure | Treat as interchangeable; optimize for speed, latency and cost per task. |
| Federate and pool | Industry data, compliance, threat signals | Partner with competitors to remove shared overhead: fraud, clinical trials, standards. |
| Own and guard | Context, customer trust, workflow logic | Never outsource the semantic map of how your business creates value. |
Seeing the forest
For fifty years the strategy textbooks taught that advantage comes from owning what others cannot. In AI, the companies with the most to own have reached the opposite conclusion. They share chips, clouds, models, agentic applications, standards and, soon, satellites. What they guard is narrower: their customers, their context and their judgment about which partner to call next.
Your competitors are already walking the forest, looking for the trees they can share. The risk in 2020 was picking the wrong partner. The risk in 2026 is believing you are large enough to stand alone.
Keep your friends close, and your frenemies on a 90-day termination clause.
Author
Jim Francis, Technology Futurist & Optimist, CEO / Founder ConceptVines
Contact: linkedin.com/in/jimgfrancis, www.conceptvines.com
Previous Redline reports: “Start in the Mailroom” (an inside-out path to deploying agents), “The Butler-Barrister” (the rise of safe personal agents), “Rembrandts in the Attic” (monetizing dormant IP and assets with AI), “Empty Cathedrals” (AI data centers without the applications) and “The Six-Month Lease” (frontier AI as a fast-depreciating asset).
Sources referenced: Reuters (via The Star) and The New Stack on SpaceX, Nvidia and Grok Bot; Data Center Dynamics, NextBigFuture, Light Reading and Tech Startups on Minihard and Starmind; The VC Corner’s SpaceX S-1 analysis and tbreak on the Colossus agreement; GeekWire on Copilot Cowork; Amazon’s OpenAI partnership announcement; MacRumors on Gemini-powered Siri; Variety on Apple’s CEO transition; The Decoder on Meta’s TPU deal; TechCrunch, 24/7 Wall St. and Macworld on Muse; Futurum on Anthropic’s IPO filing; eeNews Europe on the Agentic AI Foundation; Tech Times on the AISB federated pharma consortium; the Blue Cross Blue Shield Association report and HFMA on claims AI; Generative AI in the Newsroom on SPUR; German auto industry (VDA) software alliance coverage; U.S. Treasury report on AI and fraud data sharing.