Seven angles this week

The AI bubble narrative keeps getting harder to sustain.

New research just dropped a bottom-up, deduplicated measure of what the AI economy is actually generating on the demand side — not supply-side GPU spend, not hyperscaler capex projections, but what end customers are actually paying.

$110 billion in revenue over the past 12 months. $175 billion annualized run rate. Growing roughly three times faster than mobile or the internet did at the same stage.

Here's what I take from this as a technology executive.

The demand is real. But most enterprises — including in insurance — are still in early scaling mode. Past pilots, not yet optimized. The gap between current spend and full deployment is enormous.

That gap is actually the more important number than the headline revenue figure. It tells you how much runway is left before this becomes table stakes rather than differentiation.

The architecture decisions happening right now — which platforms, which data foundations, which vendor relationships — are going to be very difficult to reverse in 18 months. Companies treating this as a 'watch and learn' moment are not being cautious. They're falling behind a wave that isn't waiting.

The question isn't whether to invest. It's whether your current governance model is built for pilots or for scale. Most aren't. That's the real gap to close.

There's a chart I keep coming back to this week.

For 50 years — through mainframes, minicomputers, PCs, laptops, smartphones, cloud — global compute capacity grew at roughly 66% compounded annually. The trend held across every major platform shift. That's an extraordinary consistency.

In 2020, it broke. AI accelerators started coming online at a pace that shattered the 50-year line.

The last time this happened was the mid-1990s — the Pentium era, Windows 95, the early internet. That inflection lasted about a decade before reverting to the mean.

Here's why this matters for how technology leaders should be planning right now.

Most AI business cases are built on today's compute economics. But if this inflection holds — even partially — inference costs will fall faster than internal models assume. Use cases that look marginal today become viable in 18-24 months.

Real-time underwriting assistance. Continuous claims triage. Dynamic risk scoring at renewal. These aren't science fiction. They're economics problems. And the economics are moving.

The implication is straightforward: don't architect for today's cost structure. Architect for the cost structure two years from now. The companies that do will look like visionaries. The ones that don't will be rebuilding.

An audit firm just combed through $34 million in enterprise AI bills and found $1.7 million in suspected overcharges — primarily on Claude Code, across clients including Panasonic, HP, and Honda.

After disputes, about 80% of that money came back. With no admission of fault.

Anthropics's response was that there was no systematic overbilling. And they're probably right in the narrow sense. But that's not actually the problem.

The problem is structural. The same model bills differently depending on whether you buy direct, through AWS Bedrock, or through an SDK proxy. Each hop adds a meter. The meters don't agree. Retry storms rack up tokens invisibly. Failed requests sometimes get charged anyway.

This isn't fraud. It's a billing architecture that was built for the chatbot era and hasn't caught up to agent workloads.

For technology executives, especially in regulated industries like insurance, this should be a forcing function.

Token-level observability is not optional. It's a cost control requirement. Independent metering and contractual audit rights need to be in the contract before you sign — not something you negotiate after the first unexpected bill.

The labs are billing faster than they can count. Your job is to count faster than they bill.

OpenAI just taped out its first custom chip in nine months.

Industry standard is 18 to 24. OpenAI used AI to help design the chip, and then ran it on Broadcom's process. They're calling it the fastest ASIC cycle in advanced semiconductors.

The chip — Jalapeño — is already running GPT-5.3-Codex-Spark at target power. Claimed cost: 50% cheaper per inference token than current Nvidia GPUs.

Here's what I think this actually signals, beyond the hardware headline.

The chip design cycle was treated as a physical law. Two years, no exceptions. Nine months says it was a habit.

More importantly, AI just started designing the hardware it runs on. That feedback loop — intelligence accelerating the production of faster intelligence — compresses every downstream technology roadmap. Things that seemed five years away are now three. Things that seemed three years away may already be here.

For technology executives, the practical implication is this: the hardware assumptions embedded in your AI strategy have a shorter shelf life than your planning cycles assume.

If inference costs drop 50% on custom silicon in the near term, the economics of running AI at the edge — real-time fraud detection, field underwriting, live claims triage — change materially.

Any multi-year AI infrastructure contract signed today should have renegotiation triggers tied to compute cost benchmarks. The vendors already know this is coming.

Something significant happened this week that barely registered in the mainstream conversation.

OpenAI previewed its new GPT-5.6 model series to a small group of trusted partners — and the first party to see it was the US government. At the administration's request, OpenAI shared the partner list before release.

Then it turns out OpenAI, Anthropic, Google, xAI, and Microsoft were all handing over early access before a cybersecurity order was even formally signed. Meta was the lone holdout, and the government is reportedly pressing it to comply.

This is being described as voluntary. When virtually every major lab follows the same rule that doesn't legally bind them, voluntary stops meaning anything.

The release gate for frontier AI has quietly moved from the boardroom to a federal review desk. No vote, no statute, no public debate.

For technology executives in regulated industries, this is a signal worth taking seriously.

If federal oversight is arriving at the model layer, sector-specific AI regulation is not far behind. In insurance, algorithmic accountability requirements in underwriting and claims are already in motion at the NAIC and state level.

Architecture decisions being made today — model auditability, data lineage, explainability of outputs — need to be built assuming regulatory scrutiny. The time to design for compliance is before the examiner asks for documentation, not after.

AppleTalk — Apple's 1985 networking protocol — was removed from the Linux kernel last week.

Not because it was broken. Not because nobody used it. It was removed because a flood of AI-generated fix patches hit the kernel mailing list, none reviewed by anyone, and the volume made maintenance impossible.

The maintainer's decision: delete 4,000 lines and end it rather than drown in help nobody asked for.

This pattern is going to arrive inside enterprise technology stacks. Probably sooner than most architecture teams are ready for.

Insurance platforms carry decades of legacy code — policy administration systems, rating engines, claims processing modules — that has been stable for years precisely because nobody touches it. That stability was a feature. AI-assisted developers can now generate plausible patches for any module they can find, at a volume no human team can review.

This doesn't mean blocking AI from codebases. It means establishing governance before the flood arrives, not after.

Core system modules, regulatory calculation logic, and data-layer components need designated review gates. The question every architecture team should be asking right now: if AI-generated changes started hitting your oldest, most critical code, do you have the process to catch the bad ones?

AppleTalk didn't die of old age. It drowned in help nobody asked for. That's a new failure mode worth taking seriously.

Dylan Field, Figma's CEO, said something this week that I think most technology organizations are not acting on.

'AI is trained on the distribution. The most interesting, differentiated work will be out of distribution by definition.'

The context was design. But the point applies to every domain where AI is now doing the execution work.

We're investing heavily in AI tools that make coding faster, documentation faster, testing faster. That's real value. But it's also commoditizing. Every competitor has access to the same tools.

What doesn't commoditize is the judgment layer above the code. The ability to specify what to build, for whom, and why. To decide which customer interactions should be automated and which should stay human. To design agent workflows that build trust rather than erode it.

In insurance, this shows up concretely. Claims automation and straight-through processing are achievable and worthwhile. But they're table stakes — every carrier will get there. The differentiation will come from the quality of problem framing upstream.

Which data signals actually improve risk selection versus which create regulatory exposure? How do you design a renewal experience that uses AI to feel more personal, not less? Where does automation in claims handling improve customer outcomes, and where does it feel like abandonment?

Those are design and judgment problems. AI doesn't solve them. It amplifies whoever is asking the right questions.

Investing in AI tools without investing in the thinking capacity to direct them well produces faster mediocrity.