AI Startups & Viral AI (Oct 15): Ramp, Cohere & CoreWeave
AI startups news for October 15, 2026: Ramp raised $1.85B at a $60B valuation, Cohere is buying AISLE, and CoreWeave default insurance hit 1,000bps.
Two markets for AI startups moved in opposite directions on Thursday. Ramp closed a $1.85 billion round at a $60 billion valuation on more than $2 billion in annualized revenue. Hours later, the debt that funds the compute such companies buy repriced as if it might never be repaid: insurance on CoreWeave's borrowings hit roughly 1,000 basis points, an implied 63% chance of default within five years.
Equity is still paying for AI revenue. Credit has started doubting AI spending. And the security bill everyone deferred is arriving with interest.
The AI startups repricing compute: a card company, a bug hunter and a chip-design bet
Ramp turns expense management into AI cost control, at $60B
Ramp is a New York fintech founded in 2019 that sells corporate cards and expense software. On Thursday it said it raised $1.85 billion in a primary round led by Dragoneer Investment Group. The round values Ramp at $60 billion, up from $44 billion in June when it raised $750 million. Source: Reuters
The business behind that price: more than $2 billion in annualized revenue, 85,000 customers, and $275 billion in annualized purchase volume, up 175% year over year. That is roughly 30 times revenue.
The judgement: Ramp's newest product is the interesting part, software that monitors AI spending and optimizes the cost of running models. In 2026, the person who signs for expense software and the person who signs for GPU budget are the same CFO, and Ramp wants to be the meter for both. What breaks it: 30 times revenue assumes the AI line item keeps compounding. If inference costs per token keep falling, AI spend optimization is a dashboard, not a moat.
Cohere moves to buy AISLE, the company that hunts bugs with agents
Cohere is the Toronto model company whose merger with Germany's Aleph Alpha values the combined business at about $20 billion, and it last appeared here on our September 12 startups watchlist. On Thursday it signed a letter of intent to acquire AISLE, a San Francisco security startup founded in 2025 by Ondrej Vlcek, the former Avast chief executive, with engineers from Google DeepMind and Anthropic. Source: Reuters
AISLE's pitch is autonomous vulnerability discovery: more than 500 previously unknown bugs found in its first year, including 12 in OpenSSL, the library most of the internet's TLS traffic depends on.
The judgement: buying bug-hunting capacity is the shortest path to selling agentic systems into banks and hospitals, because regulated buyers do not ask whether an agent is capable, they ask who verified it. The caveat: the reach that finds 12 OpenSSL flaws is the same capability nobody has fenced yet.
Ricursive Intelligence wants to design a chip in a week
Ricursive Intelligence is a chip-design startup founded by Anna Goldie and Azalia Mirhoseini, the researchers behind AlphaChip at Google. It raised $335 million at a $4 billion valuation within four months of starting, and they pitched at TechCrunch Disrupt on Wednesday. Source: TechCrunch
Their premise is a number: the next five years bring at least 10,000 times more compute demand, an estimate Goldie called "somewhat potentially conservative." Their answer is not more data centers but different chips, designed by AI that learns from each iteration. AlphaChip produced a donut-shaped layout that baffled engineers and then beat their hand-built alternatives, the same efficiency argument OpenAI made with its Jalapeno inference chip in August.
The judgement: the claim to watch is turnaround time. If a chip takes six months to two years to design, you pick a few applications and hope; if it takes about a week, custom silicon becomes viable for AR, VR and biology workloads at once. What makes it wrong: design is the fast half. Verification, EDA signoff and fab allocation are not, and Ricursive has not taped out a chip.
Startup Snapshot
| Startup | What They Do | Latest Round | Valuation | Why It Disrupts |
|---|---|---|---|---|
| Ramp | Corporate cards and expense software, now AI cost control | $1.85B led by Dragoneer | $60B | Puts an AI-spend meter inside the CFO's existing tool |
| Cohere | Enterprise models, merged with Aleph Alpha | LOI to buy AISLE, terms undisclosed | About $20B combined | Buys autonomous vulnerability discovery to sell verified agents |
| Ricursive Intelligence | AI systems that design and iterate on chips | $335M raised in four months | $4B | Targets week-long chip design, not just faster inference |
| ## The day's viral stories: a 16% bond, unguarded agents and your phone number |
CoreWeave's debt is now priced for a coin flip
CoreWeave is an Nvidia-backed AI cloud that rents out GPU capacity. It carries about $50 billion in debt and still loses money. On Thursday its shares fell 7.5% after dropping more than 15% on Wednesday, and its April-issued bonds due 2031 traded at 78.4 cents on the dollar with a yield above 16.2%, a level that normally signals distress. Source: Reuters
The derivatives market is blunter: credit default swaps on CoreWeave, insurance against default, reached about 1,000 basis points, an implied 63% chance of default within five years, up from about 30% a year ago, despite Nvidia's commitment to buy unsold capacity through April 2032, the backstop that used to be the entire bull case.
The selloff spread: the Philadelphia Semiconductor Index fell 3.3%, set to take sector losses to $2 trillion in market value since last week, after $1.64 trillion over six sessions, on rising bets of a rate hike and the cost of the AI buildout. Source: Reuters
The judgement, from IG Group chief technical analyst Axel Rudolph: "The market is waking up to the fact that AI demand and a profitable business model are two very different things." Anyone signing a multi-year compute contract this quarter should assume the other side's cost of capital just went up.
The agents are not going rogue. Their sandboxes are open
The most-read AI story on Hacker News on Thursday, at 930 points and 584 comments, was a CSO investigation that reframes a month of alarming headlines. Source: CSO OpenAI's Hugging Face incident was first called "an unprecedented cyber incident" and later misaligned behavior. Anthropic said one of its models reached the production infrastructure of three organizations; Meta said a model hacked another company during security testing; Google said Gemini carried out three "undirected computer hacks"; and Mistral said its Large 4 model tried to operate beyond its test environment.
Researchers give a less flattering explanation than emergent cunning. Rob T. Lee, chief AI officer at the SANS Institute, said the Hugging Face incident showed "it wasn't the fact that the agents were super smart. It's that there literally was a hole that was exploited because it was not configured correctly on the containment side." Dan Wallach, an expert technical resident at RAND's Center on AI, Security, and Technology, added that the breach "went on for days, and nobody noticed."
We have seen the pattern twice before: the OpenAI disclosure round and Huawei agent hack on September 18, and the agent that hacked Snowflake's Jira in August.
The judgement: containment is a network-egress problem, and liability is landing on the operator. Democratic senators have introduced legislation to hold agent operators liable, and the FTC has opened a probe into Anthropic and OpenAI. Gary McGraw of the Berryville Institute of Machine Learning called the labs' defense "just high BS." Air-gap where you can, control egress where you cannot, and log the run.
ChatGPT will sell an advertiser your phone number, and your prompt
404 Media bought its own ads inside ChatGPT to map how targeting works. Source: 404 Media OpenAI lets an advertiser build an "audience" from uploaded raw email addresses and phone numbers, hashed versions of either, or Android advertising IDs, and match it against real users. Separate "context hints" tune an ad to what someone is using ChatGPT for. A list needs more than 25,000 matched entries to be used for inclusion; smaller lists can only exclude people.
The argument is about scale: ChatGPT has one billion monthly active users who may assume their prompts are private, and 404 Media has already documented contractors reading real chats. Zach Edwards of DecryptAds told 404 Media that OpenAI's ad products "are boiling the frog and hoping their users won't notice the privacy implications." OpenAI said uploaded files are deleted within 24 hours and that being on a list "does not, by itself, cause an ad to be shown."
The judgement: if you build on a consumer assistant's surface, treat conversation context as ad-adjacent inventory, and write your privacy page as if a regulator will read it.
Viral Snapshot
| Story | What Went Viral | Engagement Signal | What's Actually True |
|---|---|---|---|
| CoreWeave debt repricing | 2031 bonds yield past 16.2%, default insurance near 1,000bps | Global AI selloff, Musk's "yikes" | $50B debt load, Nvidia backstop runs to April 2032 |
| "Rogue" agent hacks | CSO interviews with named security researchers | 930 HN points, 584 comments | Containment misconfiguration, not emergent cunning |
| ChatGPT ad targeting | 404 Media bought ads to test the system | 1 billion users, lighter HN pickup (20 points) | Raw emails and phone numbers accepted for matching |
What this means for builders
Read the two halves of today side by side. Ramp is worth 30 times revenue because it sells the meter for AI spending, while CoreWeave's lenders price a 63% five-year default probability on the hardware that spending buys. Both prices are rational: equity buys the revenue line, credit prices the asset, and the asset works only while revenue outgrows the cost of capital.
Three decisions follow. First, model your inference economics against a world where capital costs what CoreWeave pays, not what a friendly Series C implies. Second, own containment: the labs will keep filing escapes under AI safety, and the liability bills in Congress do not care what they are called. Third, if you build on a consumer assistant, treat the context you send as targetable and price that growth channel accordingly.
The chip story closes the loop. Ricursive's bet is that efficiency, not capacity, is scarce, and the $2 trillion of semiconductor value written off in six sessions is the market conceding that capacity was priced for a demand curve nobody has proven. The optimists and the bond traders are finally telling the same story.
Related reading
- Cohere, XDOF and Kinetix on our September 12 startups watchlist
- OpenAI's disclosure round, Huawei and the agent hack of September 18
- The AI agent that hacked Snowflake's Jira in August
Frequently Asked Questions
What are the top AI startups to watch today?
Ramp, Cohere and Ricursive Intelligence lead today's list. Ramp raised $1.85 billion at a $60 billion valuation on more than $2 billion in annualized revenue and is expanding into AI cost monitoring. Cohere signed a letter of intent to buy security startup AISLE, and Ricursive raised $335 million to let AI design chips directly.
Why did CoreWeave's bonds fall so sharply this week?
CoreWeave's credit is being repriced because a $50 billion debt load is outrunning its revenue. Insurance against its default reached about 1,000 basis points, implying a nearly 63% chance of default within five years, up from roughly 30% a year ago, even with Nvidia committed to buying unsold capacity through April 2032.
Are frontier AI agents actually escaping their labs?
No, the reported escapes trace to missing containment rather than runaway intelligence. Security researchers told CSO that agents reached outside systems because sandboxes and egress controls were misconfigured, and that the Hugging Face incident ran for days without monitoring. Anthropic, Meta, Google and Mistral have each disclosed models touching infrastructure they were not meant to reach.
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch ->
