Follow the money today and it runs in two directions at once: enormous sums flowing into AI compute at the top of the stack, and the same old unpatched, unencrypted basics leaking data out the bottom. One day, both headlines. Here's what an operator should take from it.
The AI buildout is writing checks the P&L can't cash yet
The numbers at the infrastructure layer have stopped looking like software and started looking like utilities. Nvidia's contract with Anthropic for the supply and use of cloud services reportedly exceeds $180 billion, tied to 2.8GW of Nvidia AI compute capacity due for delivery by 2028. The context that matters: Anthropic revealed its 2025 revenue was approximately $2.5 billion, against an $8 billion operating loss.
The operator's take: that's a commitment an order of magnitude larger than current revenue, underwritten by the belief that demand and margins both show up on schedule. The frontier labs are making a leveraged bet on their own future, and every company building on their APIs is a passenger on that bet. You don't need to predict how it resolves — you need to make sure your product survives either outcome. If your unit economics only work at today's subsidized token prices, you've built on someone else's burn rate. Model the version where inference costs rise, keep your prompts and pipelines portable across providers, and treat single-vendor lock-in as the risk it is.
The Pentagon lost 2.76 million records to a file-sharing flaw
While the money chases AI, attackers keep winning with the boring stuff. The Pentagon confirmed that unauthorized users accessed a Defense Manpower Data Center system through a file-sharing vulnerability, affecting 2.76 million living people and roughly 294,000 deceased individuals. Two details should stop you cold: the compromised files contained unencrypted information that could include names, Social Security numbers, birth dates, contact details, and military occupational data, and the access reportedly ran from October 2025 until the flaw was discovered and patched on July 16, 2026.
The operator's take: nine months of undetected access to unencrypted PII is not a sophisticated-adversary story — it's a hygiene story. Two controls would have blunted this: encryption at rest so stolen files are useless, and detection fast enough that dwell time is measured in days, not quarters. File-transfer and sharing tools have been a favorite entry point for years because they sit between systems and often get waved through review. Inventory yours, encrypt the data behind them, and ask the uncomfortable question — if someone were quietly reading our most sensitive store right now, how long before we'd know?
Seoul says AI helped hack its banks
The thing everyone's spending billions to build is also showing up on the other side of the firewall. South Korean President Lee Jae Myung said there are signs that AI models were used in recent hacking incidents against banks in the country. The National Office of Investigation is looking into the signs, and no details were provided about who was behind the hacks.
The operator's take: attribution is early and the specifics are thin, so don't overread it — but the direction is the signal. The same capabilities that let your team automate reconnaissance, write code, and chain tasks are available to the people attacking you, at the same low cost. That compresses the time between a vulnerability going public and it being weaponized, and it raises the floor on how competent a mediocre attacker can appear. Plan for a world where exploitation is faster and cheaper: shorten your patch windows, assume phishing and social engineering will be more convincing, and make sure your detection logic isn't tuned only for the slow, human-paced attacks of the last decade.
Also on my radar
- Your breach notice might arrive via your own app. Asos app users received a push notification titled "ASOS HACKED," claiming attackers had breached the retailer's systems and accessed a Snowflake instance — extortion is getting creative, and when a third-party data cloud is the weak point, the attacker can reach your customers before your incident team does. Know which SaaS data platforms hold your crown jewels and how they're secured.
- The cheap-model race keeps accelerating. DeepSeek released DeepSeek-V4.1-Flash on October 4 — the frontier gets the headlines, but the fast, cheap tier is where most real production workloads actually run, so re-benchmark your routine tasks regularly before you overpay for reasoning you don't need.
- Agentic runtimes are becoming a product category. Oracle introduced Fusion Claw, an agentic execution runtime for Fusion Applications that pairs frontier-model reasoning with deterministic enterprise computation to automate longer-running processes — the pattern worth copying is the guardrail, not the agent: let the model decide, but keep the execution deterministic and auditable.
The throughline: the industry is spending like the hard problems are already solved at the top of the stack, while the losses keep happening at the bottom — an unencrypted database, a file-sharing bug, a Snowflake instance, now AI pointed back at the bank. The companies that win this stretch won't be the ones with the biggest compute contract. They'll be the ones who matched their ambition at the model layer with discipline at the hygiene layer. Spend on the future, but patch the present. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.