The Board Room

The pattern across six incidents

Bitget's $388M loss is the clearest template. Attackers exploited a flaw in a third-party security product and took valid admin credentials. They then sent two small test transfers that slipped under the risk threshold and began large withdrawals 30 minutes later. Risky.Biz and Daniel Miessler read it the same way: attackers didn't break the exchange itself, they broke a tool bought to protect it. Bitget's Protection Fund covers the loss. That reserve is the only reason this is a bad quarter rather than an existential one.

NetScaler repeats the pattern at the network edge. CyberScoop reports that Citrix took most of a weekend to confirm active exploitation of CVE-2026-88771 and CVE-2026-88772. Public exploit code was available, tens of thousands of instances were exposed, and CISA publicly criticized the response. SANS editors say the flaws were reportedly exploited for weeks before confirmation. Their detection guidance shapes what your team will tell you. Citrix's detection script depends on log history many teams don't keep on the device, so a team with short retention can run it and report a false all-clear. Oracle's PeopleSoft CVE-2026-35273 shows the same trap. A firewall rule meant to block it was bypassed with URL encoding, and mass exploitation returned three months after the patch shipped.

Destruction changes the metric

Microsoft now sees Jadepuffer, an autonomous attacker, destroying Azure resources. That shifts the board metric from time-to-detect to time-to-restore. CSO First Look adds the economics: at about $25 per target, hitting a thousand organizations costs roughly $25,000. Defense that scales with analyst headcount loses that race. The posture that holds limits what one compromised identity can delete and keeps backups out of that identity's reach. Microsoft is hedged on whether exposed credentials were the way in, and the number of victims is not public.

Liability now follows the vendor chain

SANS NewsBites highlights Labcorp's settlement with 44 state attorneys general: $2.3M, though GovInfoSecurity reports $2.2M. It covers a 2019 breach at AMCA, Labcorp's debt collector, which has since dissolved. Seven years later the liability landed on the client. The settlement mandates four terms that plaintiffs and customers will now cite:

• security requirements in vendor contracts
• a response plan for vendor incidents
• limits on the data shared with vendors
• a team that verifies vendor compliance

Kiteworks showed the better version of vendor behavior. It told customers to take servers offline based on federal intelligence, then found and fixed a previously unknown critical flaw during the downtime.

In Bitget, NetScaler, PeopleSoft, Jadepuffer and Labcorp, the failing controls are the ones leaders assume belong to someone else: vendor security, vendor disclosure and workaround controls such as firewall rules. That makes this a governance and contract question as much as a patching one.

Patched is not clean, mitigated is not fixed, and a vendor's balance sheet does not cap your exposure.

The loop the prospectus exposes

The Information's Martin Peers describes a chain that feeds itself. Frontier buyers demand flexibility. Suppliers' order backlogs soften. Lenders discount contracts that can be canceled. Build-outs slow. Peers sees the same constraint limiting Anthropic's own growth, which is why he calls a delay to its roughly $100B listing a question of "when, not whether." Holtec, a supplier to data centers, cited "uncertainty over data center development" when it postponed its IPO in mid-September.

Anthropic's reported numbers explain its caution. Reuters' reading of the confidential draft shows a 2025 net loss of roughly $42B. About $34B of that is a non-cash accounting charge, which leaves an operating loss above $8B on about $4.6B of revenue. Nearly a quarter of revenue came from two unnamed customers. Paul Smalera's Augment calculates that compute cost per revenue dollar fell from about $6.40 to $1.60 in a year. Augment also flags a reported $518B in obligations, set against $20.28B of cash, as the least reliable figure: it may be total commitments rather than one year's spending. Augment discloses a position in Anthropic, so weigh its framing accordingly.

Power is being financed the other way

Further down the stack, the risk moves toward the tenant. Fitch estimates that new grid-connected plants and transmission take 5–10 years. Williams brought 200 MW of on-site generation online in under 18 months, so operators pay a premium for speed. A Blackstone-led group paid $5.3B for 49% of five Williams gas projects. Investors now want 10–15 year terms plus renewals. Meanwhile, IDF's CEO says a power block is "significantly amortized" in 6–8 years. On a 15-year contract, that leaves the last seven to nine years running on a largely paid-off asset. IDF also says it can sell the power into the grid if the tenant leaves.

The capital behind all this is crowding in. Tech accounts for about 60% of U.S. convertible bond issuance this year, roughly $78B through Sept. 11, and that total excludes CoreWeave's $4.2B deal. The cost of AI-specific debt is also rising.

What this means for your contracts

That asymmetry is your leverage. Without meaning to, the best-informed buyer in the market has shown what a sophisticated compute contract looks like. Suppliers who need your credit to finance their builds have their own amortization math on the record. Together, those give you grounds to ask for flex-down rights, lower prices once the asset is paid off, and termination rights when grid power arrives.

Sources split on how to time deals with the labs. Augment argues for 24–36 month Anthropic commitments with price caps before the prospectus goes public, because its customer concentration gives Anthropic a reason to sign diversified revenue now. Others argue for holding commitments until the post-DevDay pricing settles. The defensible middle is to use the pre-IPO window to win flexibility rather than volume: price caps, capacity guarantees, notice before models are retired, and exit rights, not larger minimums.

A lab that won't hold compute risk beyond 90 days is telling you what that risk is worth.

Two data points that should change how you sign

Per-task economics are moving faster than any buying cycle. AINews reports that GPT-6.1 Sol completes tasks for about $0.72, versus $3.26 for GPT-6 Astra, while scoring one point below it on the Artificial Analysis index. Anthropic kept Sonnet 5.5 at $2/$10 per million tokens and claims up to 30% lower cost per task. Simplifying AI notes that Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%). Builder loyalty moved just as fast. Daniel Miessler reports that Claude went from trailing Codex to twice its popularity in T3 Code in about two weeks. What drove it was usage economics and harness quality, not benchmarks.

Caveats: the Sol figures come from OpenAI and one outside evaluator. AINews also notes Sol uses 10–30% more output tokens, so real savings trail the list-price math.

Why the vendors want your budget pooled now

Pricing and loyalty are that volatile, yet both leading labs are building ways to make your commitment sticky. OpenAI's Marketplace can route your OpenAI commitment to open models through Baseten. OpenAI keeps the billing relationship even when a workload leaves its models. Turing Post puts it this way: "OpenAI is willing to lose the runtime as long as it keeps the meter." Dots are always-on agents, each with its own cloud computer and 4,000+ app connectors. They extend that meter to agent labor. Specialist Dots add organizational credentials. TLDR IT reports that Anthropic's Claude marketplace, with 2,000+ connectors and partner agents, lets eligible partner software draw on Anthropic commitments. Which partners qualify, and what share of a commitment is eligible, were not disclosed.

The quieter move is routing, meaning which model handles each task. Turing Post counts a roughly 30x per-token price spread between Sol standard and Astra Ultrafast inside OpenAI's own catalog. Dots run only on Astra today. OpenAI says it is working on routing and that simplicity will win. In practice, the vendor intends to make your cost-quality trade-off for you. Flat seats are also losing value. AINews reports the new plan structure roughly halved the value of the old Pro 200 plan, which pushes spending toward metered agent usage.

Where sources agree and where they split

Every source that covered DevDay agrees that routing and evaluation belong in-house. They split on commitments. AINews advises freezing new single-vendor commitments until they include portability. Others argue that the period before Anthropic's IPO is when buyers have the most leverage to lock in terms. The two views can be reconciled: commit, but only on terms that survive a change in which lab is ahead.

If you sell software, the same hooks force a channel decision. A marketplace listing turns a new budget request into a draw on money the customer has already approved. It also lets another company's agent own the user's intent. In Casey Newton's hands-on test, a Dot did about two hours of cross-app work in 15 minutes. That shows how quickly your product can become plumbing that someone else's agent calls. Newton had the product for only a couple of hours, and pricing is unconfirmed.

When model leadership lasts weeks, the most valuable clause in an AI contract is the one that lets your spend follow the winner.