Seven angles this week

Anthropic opened its books this week, and the headline everyone chased was the wrong one.

Yes, the business is profitable. Yes, AI is 'real.' We can stop arguing about that.

The number that should hold your attention is the compute commitment: somewhere between $518 billion and $807 billion, spread across eight named counterparties. Each of those counterparties has its own debt, its own partners, its own bad week.

One company opened its books and you could see the leverage running through the walls of the whole block.

Here is what that means if you run technology for a large enterprise. When you build a core workflow on a frontier model, you are not just betting on the model. You are betting on the financing chain underneath it. The data center lender. The neocloud. The side deals none of us can see.

So the right question in your next vendor review is not 'which model scores highest.' It is 'how fast can I swap this out if the provider's balance sheet seizes up.' If the honest answer is measured in quarters, you have a continuity problem hiding behind a capability story.

Treat model providers the way you treat any critical supplier with concentrated counterparty risk. Understand who funds them. Build for substitution before you need it.

The capability is dazzling. The plumbing is fragile. Both things are true.

What is your swap time today?

The most useful AI observation I read this week came from Palantir's CTO, and it reframes almost every business case I see.

The companies trying to use AI to cut headcount are mostly failing — even at cutting headcount. The companies using AI to do things that were always beyond their reach are winning, often dramatically.

Why?

Because if your goal is to delete the human, the AI has to do 100% of what that human did. Miss 2% and you've failed. That's a brutal bar. But if your goal is to do work you could never previously afford to attempt, then every gain counts and nothing has to be perfect.

There's a second insight underneath it. The atomic unit of value is not the worker. It's the workflow. You can make a single analyst twice as fast and generate zero return, because the work still bottlenecks at the ten other steps around them.

For those of us who sit between AI and the core platforms, this is the whole game. The hard, unglamorous part is the integration layer — where a probabilistic model meets your deterministic systems, your permissions, your real-world consequences. The labs don't own that layer. We do.

So before the next pilot, ask two questions. Are we redesigning the whole workflow, or bolting AI onto one step? And are we chasing ambition, or just austerity?

The austerity cases keep disappointing. The ambition cases keep surprising. That pattern is worth taking seriously.

Here is the question I think every technology leader should be able to answer by the end of this year: if all the intelligence is compounding into model weights you don't control, what exactly is the theory of your firm?

It sounds abstract. It isn't.

When you route your most valuable operational decisions through a frontier model and hand over the data, you are quietly transferring the thing that makes you different into a commons your competitors can rent for the same price you pay.

Your edge becomes everyone's baseline.

The counter-move isn't to abandon the labs. It's to be clear-eyed about which asset actually matters. It is not the model. It's the pipeline — the operational exhaust only your business generates. Every pricing decision, every claims outcome, every underwriting judgment is training data no competitor can reproduce. That's a private environment for making models that beat the general-purpose frontier on the things you specifically care about.

And notably, this can be fast. The claim I keep hearing is that with the right data already flowing, a task-specific model takes about a week to stand up, not a two-year moonshot.

So two priorities. First, retain and organize your operational data as the strategic asset it is. Second, build the ability to swap models in and out in days, so no single provider decision becomes a one-way door.

Sovereignty here doesn't mean isolation. It means keeping what makes you, you.

Are you building that loop, or renting someone else's brain?

A frontier lab's flagship model loses to a plain, untuned model on a real business task. According to the benchmarks, that shouldn't be possible.

It happened anyway — repeatedly — in Palantir's own operations. And it should change how all of us buy AI.

The lesson: intelligence is jagged, not a single ladder you climb. A model can be brilliant on the benchmark everyone cites and mediocre on the specific, messy thing your business needs. The public leaderboard is measuring a general capability. You don't operate in general. You operate in claims, in underwriting submissions, in the weird edge cases your industry throws every day.

The phrase I'm taking into every model review from now on: bench-making, not bench-maxxing.

The right number of benchmarks isn't one global chart. It's one per company — yours. Built from your workflows, your documents, your definition of a good answer. Once you have that, the whole conversation changes. You stop chasing the newest, shiniest release and start making real trade-offs: latency versus cost versus accuracy, on criteria you actually control.

And often the surprising winner is a smaller, cheaper, self-hosted model — which cuts your cost and your dependency at the same time.

The team that builds disciplined internal evaluations holds the leverage in every vendor negotiation, because they can prove what works instead of trusting someone's slide.

If your model choice today rests on a public leaderboard, you're outsourcing a decision only you can make well.

What's on your benchmark?

Microsoft just offered voluntary exits to around 60,000 people — more than a quarter of its workforce — while posting the healthiest margins in its history.

That combination should stop you.

The easy explanations don't hold. It's not Covid bloat; that was absorbed years ago, and the offer isn't aimed at the pandemic hires. It's not desperate cost-cutting to fund data centers; this is a company printing money, with revenue per employee up two-thirds since 2022 to about $1.5 million.

A firm with record margins does not carry 60,000 unproductive passengers.

So what is it? It's confidence. Handing a quarter of your staff the door, with severance that costs real money, only makes sense if leadership genuinely believes it can run dramatically leaner next year. This is a bet on productivity they can already see coming.

The underlying signal is one I'd watch across every large company: revenue growth and headcount growth are decoupling. At Microsoft they've been diverging for four straight years, and the gap is the widest on record.

For those of us planning technology and workforce, this flips the standard question. It's no longer 'how many people do we need to grow.' It's 'how much can we grow without adding anyone.'

That's an uncomfortable reframe. But if the most profitable software company on earth is betting its future on it, your board will eventually ask you the same question.

Better to have modeled the answer before they do.

Everyone talks about agentic AI as a smarter assistant. In financial services, that framing badly undersells what's happening — and badly understates the risk.

The real change is economic. When an AI agent can shop, compare, and bind insurance coverage on a customer's behalf, it doesn't just speed up distribution. It rewrites the margin structure and moves the customer relationship at the same time. The broker channel and the direct channel both feel it. That's not a feature. That's a change to how the business makes money.

And here's the part I see teams underestimate: agents that decide and act across workflows break the governance model we built for software that only ever suggested.

When a human makes a decision, you have accountability by default. When a machine initiates the action — binds the policy, flags the transaction, moves the case forward — who owns that decision? Where's the audit trail? Do your permission boundaries hold when it's an agent, not a person, pushing the button?

My strong view: solve governance and architecture before you deploy, not after. Model your enterprise as a system that behaves identically whether a human or an agent drives each step, with clear accountability and a trace for every transition.

The firms that do this will deploy agents at scale with confidence. The firms that deploy first and govern later will spend next year cleaning up incidents and rebuilding trust.

Capability is not the bottleneck anymore. Accountability is.

Which one is your team actually working on?

Two stories ran the same week and together they tell you where AI risk is actually landing.

First, a wave of serious security vulnerabilities — PHP, Node.js, Java, Cisco, supply-chain breaches — more in one week than feels normal. The reason is structural. Software has always rested on a creaky, human-authored foundation where the real test was 'does it work,' not 'is it secure against every possible exploit.' Attackers used to be limited by human time and ingenuity. AI removes that limit. It substitutes endless iteration for cleverness, and it's getting clever too.

The open internet, where anyone could stand up a public server, is effectively over. Not because of despair — because casually exposing an app to the world is now reckless.

The constructive response is to go smaller: narrowly-scoped tools built for one team, one purpose, defended continuously by agents. Smallness is becoming the safer default, and AI finally makes it practical.

Second story: HSBC is looking to offload credit risk on data center loans to insurers.

Sit with that. The AI infrastructure boom is now arriving on insurers' desks as an underwriting line. It's a real opportunity — and a real concentration trap. You'd be pricing credit risk on assets whose value depends entirely on an AI demand curve none of us can forecast with confidence.

The lesson across both: AI isn't just a tool you adopt. It's reshaping the risk landscape you operate in — in your own systems and on your balance sheet.

Are you watching both?