Three infrastructure announcements on November 5 sit at different layers of the same stack. A funding round for querying huge datasets with AI agents, a networking deal for a new inference cloud, and a storage test at CERN for keeping data for decades.

Ocient: 100 million dollars and revenue set to triple

Ocient, a Chicago data analytics company, raised a $100 million Series C co-led by BOKA Capital and OCA Ventures. Returning investors include Greycroft, In-Q-Tel, Blue Bear Capital, Buoyant Ventures, PSP Growth and Pritzker Group Venture Capital. Total funding is now $260 million. No valuation was given.

OpenAI released GPT-6 Cyber on November 5, its fourth security model this year and the first it rates “Critical” for cybersecurity under its own Preparedness Framework. Researchers using it found 81 previously unknown high-severity flaws in Firefox. Mozilla has started shipping fixes. Against the Linux kernel, the model reproduced 1,724 known vulnerabilities and built working exploits for 948 of them. The previous model, GPT-5.6 Sol, managed 166 in the same test.

Most access control programs are built around getting people in. New starters get a laptop, an identity, a set of groups and a list of applications on day one, usually through a checklist that HR and IT have tuned for years. Getting people out gets far less attention. The result is predictable: former employees, contractors and interns who still have working access weeks or months after they left, to systems nobody remembered they had.

I like maintainers. I’ve spent years reading their mailing lists, filing their bugs and depending on code they wrote at night for nothing, and I run a few small projects myself. So I’m the last person who wants to argue that the human at the centre of an open source project matters less than we say. My claim is narrower than that. It’s about what happens on the days that person isn’t there.

Hyper Crux: Round 7 is finished. V1, T5 and F5 are merged on main, all 76 planted bugs were caught, and both workflows passed on GitHub. The round used about 1.7 million tokens over about an hour and a half, including the timed benchmark run.

V1, vector arrays and Nearest: each table’s vectors are stored in memory, and Nearest searches them exactly. In rough timings on the shared machine, all three searches beat their targets:

Anyone who has built the software side of a connected product knows the glue job. The device keeps its readings in one format and the gateway wants rows in a SQL database, so somebody writes a converter. Then somebody has to keep that converter alive for as long as the device stays in the field, which can easily be ten years.

AltSql takes the converter out. Its base engine, AltSql Core, compiles to about 15 KB of code for a microcontroller. The device keeps its readings and settings in its own flash and goes on deciding from its own data when the connection goes. Its gateway stores the very same records, byte for byte, and answers SQL over them.

Most sites use a thin slice of nginx: TLS, a folder of static files, a couple of routes to an app, health checks and the odd reload. The config grows anyway, one regex at a time, until nobody is quite sure which block handles a given URL.

BareProxy keeps its core to that slice and makes it answer questions. bareproxy explain takes a URL and shows which rule matched, which file or backend would serve it, and why the rules above it didn’t match. bareproxy why tells the story of a request that already happened, starting from the ID it carried in a response header.

Most dashboards ask the same few questions all day long: requests per endpoint, say, or this month’s usage for one customer. The usual setup recounts raw rows on every refresh, or ships every row to a hosted service that bills by the gigabyte.

Precomputing does the counting once, as the data comes in. You write a short policy that names the answers you want and says how long each level of detail should live. It compiles to plain SQLite triggers, so every INSERT keeps those answers current and reading one is a lookup. Old detail fades on your schedule; unusual events are kept whole.

Cloud coding agents start each task on a fresh machine. Before the agent can run a single test, something has to install the right runtime and start the database, and every platform wants those instructions in its own file. Copilot runs a setup workflow; Cursor builds a Dockerfile. A team that uses two or three agents writes the same setup two or three times, and a mistake usually shows up later as an agent session that goes nowhere.

Sandisk spent $4.5 billion on its own stock in the roughly nine weeks between the board’s April 30 authorization and the July 3 fiscal year-end. On August 5 the board added another $14 billion. That leaves $15.5 billion of firepower pointed at a stock trading around $1,600, about a third below the June 25 record close of $2,354.39. The correction took the price down. It didn’t touch the cash.

The $14 billion headline matters less than the $5.035 billion of adjusted free cash flow Sandisk generated in fiscal Q4, of which the $4.5 billion repurchase absorbed 89%. Annualize that quarter and the company throws off more than $20 billion a year. With about 149 million shares outstanding, the market cap sits near $240 billion, so the stock trades at roughly an 8% free cash flow yield on the current run rate. The entire remaining authorization is less than one year of cash at the Q4 pace. At $1,600 it retires about 9.7 million shares, or around 6.5% of the company, and management’s Investor Day framework commits to returning 100% of excess cash from here. This authorization will need topping up again.