Washington Picks Sides, Silver Lake Bets $48B on Enterprise AI

So, What Actually Happened?

Friday I read a draft letter and put the coffee down. Washington is about to tell 35 partner countries they must pick one AI bloc, its own or Beijing's, and hold membership in nothing that competes. We scanned 190,000 articles this week so you don't have to. Same afternoon, Silver Lake opened talks on Workday at $48 billion, a deal that only closes if a room full of co-investors signs up to the idea that AI grows enterprise software instead of eating it. And a study went the other way entirely, contradicting the labs' autonomy claim that AI can now run its own research. Three rooms, one request. Sign here, before anyone can price what you are signing.

The Bottom Line: The ask changed this week. Nobody is being sold a capability anymore, they are being asked to underwrite one.

The Tracks That Matter

1. Washington Asks 35 Countries To Pick A Side, In Writing

The State Department is drafting a letter telling signatories of its AI Opportunity Statement to choose between Pax Silica and Beijing's rival body. The language is unusually blunt for diplomatic paper: to be part of everything is to be part of nothing. About two dozen countries have joined Pax Silica, including Japan, Australia and South Korea, and Kazakhstan is currently the only one sitting in both rooms, which is the position the letter exists to end. Beijing is not waiting either, with two central authorities jointly launching a 2026 model-and-data initiative to pull Chinese industry onto one stack. Whose chips and whose models used to be a procurement question. It is turning into a passport question, and this one has no dual-citizenship option.

Here's what works: Map every AI layer in your stack to the jurisdiction whose rules govern it. Any layer with only one is policy exposure, not vendor choice.

2. Silver Lake Wants $48 Billion Of Faith In Enterprise Software

Silver Lake is in talks to take Workday private at $48 billion, and the arithmetic is the interesting part. At $227 a share it is a 30% premium, a $53.8 billion enterprise value, and roughly $38 billion of equity, which means co-investors. Those co-investors are being asked to bet against the exact fear that nearly halved software's median revenue multiple in a year: that seat pricing dies, that customers vibe-code their own tools, that agents replace the humans you were billing for. Software buyouts sit at a decade low. Meanwhile a top analyst named Snowflake and CrowdStrike the AI software to own, which is the same trade from the other side: buy the plumbing on the public market, take the seat business private and quiet.

Here's what works: Before your next SaaS renewal, ask what the vendor charges for when headcount stops being the unit. If they cannot answer, price the risk yourself.

3. A Study Says AI Cannot Yet Run Its Own Research

A new study contradicts the claim that autonomous AI research is within reach, the claim the leading labs use to justify the next round of spending. The timing is what makes it awkward. A day earlier, the same outlet noted that milestones those researchers predicted have already fallen, which is the argument for the other side. Both hold, and that is the useful part: models keep clearing individual benchmarks while failing at the unglamorous thing that actually constitutes research, reproducing a result end to end. One lab is now training models specifically to replicate published work, which tells you exactly where the gap sits. If your 2027 roadmap assumes self-improving systems, you bought the press release rather than the paper.

Here's what works: Ask any vendor selling autonomous agents for one reproduced result, not a benchmark score. Reproduction separates a demo from a capability.

Quick hits:

• Data centers are now being designed for orbit. Insurers have entered exploratory talks on orbital data centers, with underwriters asking the only question that matters, whether the risk can be modelled at all.
• Ryanair put its whole workforce on a five-year AI contract. The airline moved 35,000 employees onto Google Cloud through 2031, which is what adoption looks like once it stops being a pilot.
• Riot landed a $9.1 billion data center deal in Texas. The Rockdale campus was contracted for $9.1 billion, converting mining capacity into AI capacity and repricing power assumptions across the region.

Signal vs. Noise

🟢 Signal: Risk assessment. Fewer pieces named it on Friday and it carried more weight in every one: insurers deciding whether orbital compute can be underwritten, a regulator proposing to charge for data-centre grid queue positions, private equity pricing an AI thesis into a $48 billion offer. Risk work moved off the compliance slide and into the deal terms. Most coverage is still grading models.

🔴 Noise: ”Agentic AI.” It pulled heavy volume again and lost ground on both counts, said far more often while less and less actually hangs off it. Nearly all of Friday's was vendor positioning: contact-center agents, marketing agents, employee-experience agents, no new named buyer attached to any of them. When the label does all the work, the label is the product.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

A research firm named retrieval architecture the backbone of generative AI customer search, a security body told teams the first question to ask about machine identities, and Cisco closed a record quarter on AI networking demand reaching industrial buyers.

Three desks filed those separately. The customer-experience desk covered the Metrigy report naming retrieval architecture as central to technical success. The security desk published the machine-identity question every team now has to answer. The equity desk wrote up Cisco's record quarter and the industrial demand behind it. Read them on one morning and none of the three is about a model. Retrieval, identity and networking sit underneath whichever model you picked, and all three got named as the thing that decides whether it works.

That is a purchasing shift, not a technical one. The budget line that said ”AI” in January is quietly splitting into a plumbing line and a licence line, and the plumbing line is the one growing.

What changes on Monday is small. Pull your last three AI invoices and separate what paid for a model from what paid for the retrieval, identity and network work around it. If you cannot split them, you have already lost the strongest argument you had going into the renewal.

By The Numbers

• Silver Lake is in talks to take Workday private at $48 billion — at $227 a share that is a 30% premium and a $53.8 billion enterprise value, struck in a year when software buyouts hit a decade low.
• The US drafted a letter to 35 AI Opportunity Statement signatories — about two dozen countries have joined Pax Silica so far, and Kazakhstan is the only one currently signed into both blocs.
• Riot landed a $9.1 billion AI data center deal in Rockdale, Texas — mining capacity converting to AI capacity at a scale that resets every power assumption in the state.
• Squadron Energy secured AU$2.7 billion to refinance hybrid assets — generation and storage financed as one asset class, which is what happens when buyers start contracting for firm power rather than cheap power.
• The Texas Space Commission awarded Axiom up to $5.5 million for orbital data-center work — small money, but the grant runs to November 2027, which is a public deadline you can hold them to.
• See what's rising across AI and data this week →

Deep Dive: The Promoter Wants An Exclusivity Clause

There was a promoter in Antwerp who handed you the season contract in March. Residency, decent fee, one clause: you do not play the other room across the square. He handed it over before the lineup was announced, before anyone knew which room would matter that year. You were signing a bet on his booking, not on your own set.

The letter says the quiet part
The State Department's draft is that clause, written by a government. Sign Pax Silica, and do not hold membership in duplicative initiatives whose expectations conflict with ours. Two dozen countries are in. One of them is in both rooms, and the letter exists to make that impossible.

The check needs believers
The Workday approach is the same clause pointed at capital. Thirty-eight billion in equity means co-investors, and co-investors are being asked to sign a view on what AI does to enterprise software before anyone has five years of evidence either way. That is not diligence. That is faith with a term sheet attached.

The underwriters said not yet
Space insurers, asked to cover orbital data centers, gave the only honest answer available: they are still working out whether the risk can be modelled. An underwriter who cannot model a risk does not price it high, they decline it. Two rooms this week issued a signature request. The third handed back a polite no, which is the most informative answer of the three.

What Actually Works

1. Name the jurisdiction on every layer: for model, compute, data store and network, write down whose rules govern it. Single-jurisdiction layers are policy risk now, not vendor risk.
2. Write down what you are guaranteeing: every AI contract signed this year carries an implicit forecast. Put it in one sentence and see whether you would defend it out loud.
3. Split the plumbing line from the licence line: retrieval, identity and network spend is growing faster than model spend. Budget them apart or renegotiate blind.
4. Demand a reproduced result: for anything sold as autonomous or self-improving, one end-to-end reproduction beats any benchmark table you will be shown.

The promoter always has a lineup in mind when he hands you the contract. Nobody signing this week has seen theirs.

What's Coming

The Grid Starts Charging For Intentions

Ofgem is proposing a commitment fee for data-centre connections, which turns a queue position into a priced option. Speculative requests have clogged that queue for two years. Once a fee lands, your announced data-centre plan and your actual one have to match.

A Named Date On The Crisis Talk

Brian Armstrong put a one-to-two-year window on a major AI crisis. Predictions are cheap; dated ones are not. When operators start timestamping the downside, boards and insurers ask for the same in writing shortly after.

Clinical AI Gets Its Audit

Nature ran a critical look at AI in medicine. Medical AI has had a long grace period on evidence quality. Journals turning the lens back on the field is usually what happens just before regulators do.

For Your Team

Monday's meeting prompt: ”For every AI system we run, whose country's rules govern it, and what exactly are we guaranteeing to the vendor who sold it to us?”

Share-worthy stat: Taking Workday private at $48 billion needs roughly $38 billion of equity. That is not a financing question, it is a poll with money attached: does AI grow enterprise software or dissolve it, and somebody has to answer this quarter.

Go deeper: Track where AI policy and AI spend are separating →

The Track of the Day

”To be part of everything is to be part of nothing.”
From the US State Department's draft letter to AI Opportunity Statement signatories

Every DJ has turned down a residency that paid well because the clause underneath it cost more than the fee. You usually find out which one it was about two years later.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: August 15, 2026 | Curated by Yves Mulkers @ Ins7ghts