The Board Room

Two ways to pay for an AI position

The Information reports ByteDance's first-half 2026 revenue grew about 30% to $120 billion while net profit fell to $20 billion under AI spending. At FY2025's implied ~21% net margin, $120 billion of first-half revenue should have produced roughly $25 billion of profit. The gap is about 430 basis points of margin, roughly $10 billion annualized, and it is the on-P&L price of an AI position. A private company with every incentive to protect earnings ahead of employee tenders and secondary pricing paid that price voluntarily. Growth did not stall to fund it. It accelerated, from 29% in 2025 to about 30%.

One caveat before anyone models this: the figures are unaudited, rounded, sourced from three anonymous individuals, and the profit decline is characterized only as a "single-digit percentage." Treat the direction as sound and the decimals as noise.

Oracle bought the same position with a different instrument. $90–95 billion of capital expenditure this year, against $125 billion of debt and a credit rating below its hyperscaler rivals, per Pivot 5. Headcount is down roughly 21,000, or 13%, in the fiscal year ended May 31, and the 2026 restructuring charge was revised up $700 million to $2.8 billion, largely severance. Atlassian cut about 1,600 in March, explicitly to self-fund AI. Autodesk cut about 1,000 in January. Atlassian and Autodesk are the mid-cap comparables boards will cite.

Where the sources agree, and where they split

Four separate reports carried the ByteDance profit figure this cycle and reached the same conclusion: AI cost has crossed from the balance sheet to the income statement. Applied AI adds the pairing operators would rather not see, margins compressing at the operators while pure-play specialists inflate, with Shield AI reportedly in talks at $20 billion-plus.

The split is over whether payroll is financing at all. Computerworld surfaces Gartner's finding that at least one in three AI-eliminated roles will be restored by 2029, at higher cost, with the warning that AI deployed primarily as a cost-cutting lever backfires in enterprises. On that reading, headcount-funded capex is a provisioned liability. The savings go back out with a rehiring premium, plus a credibility cost inside the organization that heard the original story. This quarter's severance line becomes 2029's recruiting line.

Where the money comes from decides the cost

The cheapest capacity this quarter comes from engineering unit cost down, not from new dollars of any kind. Debt and severance both buy compute, and only one of them leaves delivery capability intact. Oracle attempting its largest buildout ever with 13% fewer people is a coin flip on execution, and that coin flip is a hiring window for everyone else: three named self-funders have put senior infrastructure, delivery and platform talent into the market at once.

Margin compression is being repriced as a credential. Oracle, Atlassian and Autodesk can each point to what the compression bought. Most buyers cannot name the position, and they will be recruiting in the window those three just opened.

The default asymmetry is the disclosure

OpenAI pays hundreds of contractors under a program codenamed Project Lily to read real ChatGPT conversations, and the setting that feeds them is on by default for Free, Plus and Pro but off for Enterprise, Business and Edu, per Techpresso. A vendor whose settings protect enterprise accounts and expose consumer ones has told you where it believes the liability sits. Two details make this worse than a routine privacy story. OpenAI's Privacy Filter — the model that strips personal details before human review — concedes in its own documentation that it can miss uncommon identifiers and under-redact when context is limited. And reviewers see a user-memory summary above the prompt that can reveal prior usage and rough geography, per Pivot 5. The redaction is imperfect, and the surrounding context is richer than the prompt itself.

Then the deletion gap

In ChatGPT, deleting a chat does not delete the memory derived from it. That has to be corrected separately in Manage Memories, and project-only memory can be set only at project creation and cannot be reversed, per Simplifying AI. Employees reasonably believe deleting the chat deletes the data. If any team is running Projects on customer, candidate, employee or patient material — and someone is — your end-to-end deletion capability is weaker than your published policy claims. That is a live data-subject-rights exposure under GDPR and CCPA that costs days to close and reputation to discover late.

Then retention becomes discovery

The 3M matter flips AI governance from input control to output retention, per CSO First Look. Enterprises spent three years worrying about what employees put into AI tools; the discoverable record those interactions leave behind is the exposure nobody wrote a policy for. The worst case is compelled production of unbounded prompt history, and the earliest detection signal is a subpoena — litigation exposure arrives without a compliance deadline to plan against.

And the adversary is already inside this loop

Threat actors have stopped merely using AI and started stealing yours: pilfered AI credentials running distillation attacks — querying a model repeatedly to copy its behavior — plus automated exploitation pipelines inside victim cloud tenants. The first visible signal is an unexplained inference cost and GPU utilization spike. That telemetry already exists in your billing system; it is simply routed to Finance instead of security triage. This is the cheapest high-yield control available to you this quarter, and it requires a dotted line rather than headcount.

What sets your standard of care

Palantir and Nvidia are reportedly curbing internal AI model use over data-handling fears, per Techpresso. Two of the most technically sophisticated buyers on earth restricting their own usage establishes the bar your customers' security reviewers will apply to you within two quarters.

There is an offensive read, and it is the more valuable one. System-enforced provenance and reversible, auditable, provably-deletable memory are product features the frontier labs are visibly not prioritizing. If you sell into regulated buyers, that is an open flank — a claim you can make and verify while competitors stall in security review. Making it credibly requires the audit first: you cannot publish a deletion guarantee you have not tested against derived memory.

Your employees' prompts are now discoverable records held by a third party, and the opt-out only works going forward — which makes this a legal-hold question, not an IT setting.

The ceiling: three platforms, one package shape

The convergence matters more than any individual launch. The three largest Western consumer platforms have each independently concluded that AI belongs as a metered increment inside an existing bundle — sitting next to storage, media and device support — rather than as a product with its own price discovery. Meta One, which bundles extra AI usage with social-app tools, is the newest instance, per The Information. A reference price set by companies who can afford to sell AI at zero margin as a retention tool is a hostile environment for anyone who needs AI to carry its own P&L.

And the comfortable story that subscriptions offset AI capex is unproven at every scale point in the data, per The Information.

The floor: cut by the labs' own distribution partners

GitHub's HydraFusion research preview dispatches Copilot tasks across rival models and claims 65–67% cost reduction against a Claude Opus 5 baseline — GitHub's own number as reported by Pivot 5, not an independent benchmark. That is Microsoft's distribution arm publicly demonstrating the premium model is optional. Cognition's SWE-2, built on the open-weight Kimi K3, beats Grok 4.6 at half the cost, per ben's bites, which also reports Bolt Forge giving GLM, DeepSeek and Kimi away free until October 14 — a promotional window, not a durable price.

Google supplied the third cut by shipping Deep Research — multi-step research and synthesis — into Gemini Live's voice mode for all users including the free tier, with paid tiers buying only higher daily limits and a better generator, per Simplifying AI. If a SKU in your portfolio prices research or source aggregation as a differentiator, its floor was set to zero by a company with no interest in that revenue line.

Where the reads diverge

Both sources agree the model layer is substitutable in both directions: Salesforce shipped Koa on Nvidia's open-source Nemotron, bypassing the frontier labs, while developers run Claude Code against non-Anthropic models — Anthropic's own tooling generating inference revenue for competitors.

They disagree on where defensibility lands. The Information argues the only durable lock-in is accumulated context: universal free tiers plus five credible substitutes cap price and erase switching cost, so memory is the moat. Simplifying AI argues distribution is being rewired underneath that claim — ElevenLabs shipped voice, music, image and video into ChatGPT, Claude, Cursor, Grok and Hermes on one OAuth sign-in, no server, no API keys. Both cannot fully hold. If the connector becomes the unit of distribution, the platform owns the surface where memory accumulates, and your retention asset accrues to someone else's account.

The move

This is a repricing problem, not a rebuild. Move every price justification onto the four dimensions a free consumer feature structurally cannot reach: proprietary data access, citation auditability, compliance posture, and workflow embedding. Then measure the thing that decides whether any of it works — margin per paying AI user, not aggregate AI spend. Do it before renewals, because the alternative is a customer doing the arithmetic first and opening a conversation you did not schedule.

The winning question is no longer what you charge for AI — it is what AI makes un-cancellable, and a bundled free tier can answer that for your customer before you do.