Ramp says there was a week when its employees spent $1.5 million on AI usage fees. Most companies would have called a meeting. Ramp studied what its people were doing, built an internal system that captured the workflows that worked, and on October 13 turned that system into a product called Glass. Two days later it announced $1.85 billion of new money at a $60 billion valuation.
Look at what the product is. Glass is a record of how a company does its work, meaning which instructions and which pieces of context made a task go well, so the next employee does not start from zero. No AI lab can ship that, because no AI lab was in the room.
I think that is the right way to read the week. Almost everything announced between October 12 and October 18 was one company absorbing something that last month looked like a company of its own.
The decision that became a setting
Take the smallest example. A decision model is an AI that never writes anything. You hand it a situation and a list of options, and it returns a probability for each option, fast and cheap. Send this customer to billing or to support. Is this message a scam. Which tool should the agent use next.
On October 13 the benchmarking firm Artificial Analysis published its first independent ranking of these models. A startup called TypeSafe came out on top with a model named Jev, scoring 58, ahead of Fastino and Microsoft at 54 each. Two days later, on October 15, OpenAI made decisions a built-in part of its live voice product. A developer building a voice agent can now hand a choice to OpenAI's own decision tool, or trigger one from the transcript while the conversation is still going. The next day OpenAI made it a standard API feature to create a custom voice from a text description, which retired another small startup category in a release note. Our October 4 issue had already noted that Cloudflare copied the decision idea within days of Jev appearing.
Artificial Analysis also found a weakness that explains where the money will be. These models are overconfident on questions of knowledge and reasoning. When one says it is 80 percent sure, it is right less often than that. A fix for that gap is a business, but only for a company that holds its own history of which decisions turned out well. OpenAI can supply the tool. It cannot supply a lender's record of which approvals later defaulted.
So the tool is a feature. The record of what happened after each decision is the company.
Pressure from below
The open side pushed in from the other direction. On October 17 DeepSeek released a new version of its Flash model with open weights, which means anyone can download the trained model and run it, under the MIT license, one of the most permissive there is. DeepSeek reports that its score on Terminal-Bench 4.0, a test of how well a model works at a command line, rose from 31.2 to 42.3. That number comes from DeepSeek. As of Sunday no independent group had tested this version; Artificial Analysis's measurements cover the September release. Prices did not change: $0.30 per million input tokens (a token is roughly three quarters of a word) and $1.20 per million output tokens at peak hours. The same day DeepSeek opened search and page-fetching tools at $0.75 per thousand requests. Another small API business just moved onto the model bill.
The day before, Reflection AI released the weights of Beam under the Apache 2.0 license. Beam has 501 billion parameters, the internal numbers a model learns, with 23 billion active on any given word. It scores 30 on Artificial Analysis's Intelligence Index, second among American open models, and Amazon had a deployment path ready the same day. Its model card reportedly allows distillation, which means using a big model to teach a smaller one. That matters to any startup that wants its own cheap specialized model without a legal question hanging over it. Beam still trails the best Chinese open models on the index, so its appeal to American companies is a clean license and a familiar supplier.
Meanwhile Meta, which reorganized its AI effort only about fourteen months ago, released Muse Wonder on October 12. Artificial Analysis scored it 52, and the bigger Wonder Full 56, up from 48 for Meta's previous model six weeks earlier. Only Anthropic's Claude Opus 5.5 scores higher, at 58. Meta kept Wonder Full off its API and gave it only to subscribers of its own coding product. A lab that holds back its best model for its own coding tool is now a competitor of the startups that would have built on it.
Price lists mislead here. The number that matters is what it costs to finish a task. On Artificial Analysis's measurements, Opus 5.5's five-point lead over OpenAI's GPT-6 Astra costs 83 percent more per task, $5.98 against $3.26. Muse Wonder costs $4.52 per task, 2.8 times Meta's previous model, for four more points. In September's figures a small open model, GLM-5.3 Flash, scored 42 at about 25 cents a task, roughly a twenty-fourth of Opus's cost. Most work does not need the last five points, and anyone sending every task to the top model is overpaying.
What the labs found out about their own machines
Anthropic published three results this week that read like a lab surprising itself.
The cleanest came on October 16. Anthropic said an internal Claude research model worked out, in under half an hour, where the unexplained starting numbers behind several widely used government-standard encryption curves came from. They were produced by an old random number generator seeded with the time of day. Anthropic published the derivation and the code that regenerates the numbers, and said plainly that this is not a break in the encryption. I find this the most convincing AI science claim in a while, because anyone can run the code and check. The same day Anthropic reported that Claude improved published designs for eight quantum circuits, in some cases by an order of magnitude, scored by an automated evaluator. Collaborators checked the work; outside reproduction is still to come.
The uncomfortable one came on October 12. Anthropic's red team, the group paid to attack its own systems, reported that models of roughly 30 billion parameters or fewer can already do many of the steps needed to copy themselves onto cheap internet-connected hardware. GLM-5.2 completed a simplified version end to end in most trials, and the team ran a Gemma-class model across a cluster of Android TV boxes costing about $55 each. They say they have not shown a fully autonomous replicator, and they project that small models could cross their threshold by the end of 2026, which is the least firm part of the report. Still, the idea that dangerous capability needs rare chips looks weaker than it did. Companies that control which machine can reach what, through identity, access and network limits, have a better week than companies that sell raw compute.
The third result is Anthropic studying itself. Junior technical staff fell from 21 percent to 14 percent of its technical workforce since early 2025, and junior hiring fell from 27 percent to 9 percent. By August, hours of AI agent work ran about six to one against employee hours, and code commits per engineer were about five times what they were six months earlier. One company with the best tools in the world is a poor sample of the economy, so read it as a preview of one company's future and ignore it as an average.
The number I would watch is smaller. The share of work sessions spent coordinating, checking and steering the AI rose from about 10 percent in 2025 to about 22 percent by spring. When agents do more of the work, people spend more of their time deciding what to ask, reviewing, and handling exceptions. Glass is a product aimed at exactly that.
Two prices of money
Money also got more expensive. On October 14 the Bureau of Labor Statistics reported that September consumer prices rose 0.6 percent in the month and 3.7 percent over the year, up from 3.4 percent. Gasoline caused more than half of the monthly rise. Core prices, which leave out food and energy, rose 2.6 percent over the year, and both readings came in 0.1 point above forecasts according to CNBC. The 10-year Treasury yield, the benchmark for long-term borrowing costs, touched about 5.3 percent during the day. The Federal Reserve meets October 27 and 28, and Trading Economics puts the odds of a rate increase near even.
Higher rates hurt assets whose payoffs sit far in the future, and little in the market sits further out than a frontier AI lab. That is the context for the week's most important price signal. Anthropic held an investor day on October 14, ahead of a possible stock market listing. Reuters reported that afterward investors discussed a valuation of around $1.5 trillion, down from the roughly $2 trillion that had been floated. The reason given was slowing growth. Anthropic told investors that third-quarter revenue rose 45 percent from the second quarter to $16.8 billion, after the second quarter had more than doubled to $11.6 billion. Those are company figures shown to investors, not audited numbers, and I could not confirm that a registration statement has been publicly filed, though there have been reports of the confidential filing. Even at $1.5 trillion the price would be about 1.55 times the $965 billion valuation from May.
Ramp is the contrast. Its valuation rose about 36 percent since June, from $44 billion to $60 billion, while its annualized revenue (this month's revenue multiplied out to a full year) rose at least 33 percent, from over $1.5 billion to over $2 billion. It reported positive cash flow, more than 85,000 customers and more than $275 billion of annualized purchase volume. The price followed the business. Modal Labs, which runs other companies' AI models for them, raised $750 million at $15.75 billion with a run rate above $1 billion, about 16 times. It looks like value is drifting toward whoever runs the machinery and whoever owns the workflow, and away from the company that builds the model. The IMF's October financial stability report added a sober note, naming leverage and concentrated exposure as ways a drop in AI prices could be amplified.
Software ships on a Tuesday. Power lines do not.
The Financial Times reported on October 18 that Microsoft pulled out of a planned data center deal with Nscale in Essex, England, because the site's grid connection would not arrive until after 2030. The site was about 50 megawatts and had been tied to a planned supercomputer of 23,000 Nvidia GB300 chips. Nscale is reportedly looking for other power. Anyone who has played a strategy game knows the feeling of queuing ten thousand units and watching nothing move because the power plant is not built.
That makes the week's other physical news easier to read. Salience Labs raised a $125 million Series B led by Bullhound, with AMD's and Intel's venture arms joining, to build light-based switches that connect chips inside data centers. Ubitium raised a $14 million seed led by Ibex for a single processor aimed at robots and machines, and says its first silicon works. Mueon raised a $65 million Series A for modular computing hardware. As the chips themselves get easier to find, the wiring and the electricity decide how much of them anyone can use.
Nvidia's next system, Vera Rubin, shows the same shift. Artificial Analysis measured about 5.7 times the output per megawatt of the current GB300 system when serving DeepSeek V4 Pro to users who want quick responses, about 2.8 times on Kimi K3, and only 1.14 times at peak bulk throughput. A megawatt is how a data center buys power, so output per megawatt shows what the electricity bill turns into. The gain is large for interactive agents and small for batch work. My inference is that loans secured by older racks, priced before these numbers existed, look worse today.
What this means for founders
Imagine the sentence a larger company would write in a release note to describe your product. If it fits in one line, you have a feature, and you should know that before your next fundraise. A few questions to put to yourself this week:
- What do you still own the day OpenAI, Anthropic, Google, Meta or DeepSeek ships your core capability inside its platform?
- What does one finished customer task cost you at the reasoning effort you run in production, counting retries, tool calls and human review?
- What outcome data does your product collect that no one can rebuild from public training data?
- Can every consequential action your agent takes be traced, attributed and reversed?
- Which dependency in your business has the longest real lead time, whether data access, permits, grid power, hardware or a model you do not control?
Wedges that look better after this week are the ones with a record behind them: decision systems tied to real outcomes, agent control and audit tools, workflow memory inside systems customers already use, independent testing and verification, and genuinely scarce physical infrastructure such as interconnect and efficient chips for robots. Wedges that look worse are standalone decision APIs, voice generation, web search wrappers, general AI workspaces without distribution, and coding agents that depend on a single lab's API.
What this means for Team Ignite LPs and other investors
Our early-stage thesis gets cheaper to execute every time a capable model becomes free or nearly free, and it gets harder to defend whenever a founder's edge was access to a model. Seed checks should follow the record, the distribution and the physical bottleneck. At late stage the standard runs the other way. A large private valuation deserves the disclosure Ramp gave, covering revenue quality, cash flow and commitments, and it deserves extra scrutiny where the company's capacity depends on grid dates, cancellation terms and borrowed money. Private marks on frontier AI names are likely to drift toward whatever a public listing sets, and the discount rate will do that work before any company news does.
For other venture investors, the mispricings to look for are three. Middleware priced for today's technical necessity, when a platform can absorb it in one release. Customer outcome data, which is dull next to a benchmark chart and lasts longer. And physical AI infrastructure, which splits into scarce enabling technology and plain capacity financed on optimistic delivery dates. The first can earn a moat. The second is leveraged exposure to a commodity.
A feature is whatever a larger company can ship in a release note. A company is what takes a decade of customer history, or a utility's permission, to copy.
Sources: Last Week Ignite archive; Reflection, Beam is now available; Artificial Analysis, Muse Wonder; Artificial Analysis, Decision Index; Artificial Analysis changelog; AWS Machine Learning blog on Beam; CNBC on September CPI; Crunchbase News weekly rounds; FinanceFeeds on the Anthropic valuation.