AI Startups & Viral AI (Oct 18): Nscale, CoreWeave & DeepSeek
AI startups hit a credit wall on October 18, 2026: Microsoft quits the Nscale UK data center, Anthropic pays 10%+ for chips, CoreWeave bonds yield 15%.
Two things got more expensive for AI startups this week, and neither is a model. Compute credit repriced: CoreWeave's June bonds now yield 15%, and Anthropic is offering Blackstone over 10% to fund a chip lease. Permission repriced the same way: JPMorgan Chase, Charles Schwab and Delta are shutting agents out, and Meta deleted a scraping clause it had buried in its commercial terms. Microsoft's exit from an Nscale data center makes the first trend concrete, because the binding constraint on AI capacity is now grid power and counterparty risk rather than GPU supply. DeepSeek's new search and fetch APIs put a price on the other half of the problem: $0.75 per 1,000 requests to reach the web your agents need.
AI startups hit the credit wall: Nscale, CoreWeave and Anthropic
Microsoft walks away from Nscale's Loughton data center
Microsoft, one of Nscale's two largest customers, recently terminated its agreement for Nscale's Loughton, Essex data center. Source: Reuters Neither side explained the exit, and Microsoft declined to comment.
The Telegraph's reporting points at power. Grid officials told Nscale they did not expect a supply for Loughton until the 2030s, after repeated project delays. Nscale calls it "this unfortunate London-wide grid issue" and says it is exploring alternative power sources and other UK sites.
Nscale is an infrastructure bellwether here: Nvidia-backed, competing with CoreWeave, Nebius, Crusoe and Lambda, and scaling from $100 million to more than $103 billion in contracted value in two and a half years. That curve sits next to our September look at Nscale's pipeline when Anthropic and Microsoft were both signing as its IPO customers.
Judgement: a contracted-value number is only as good as the power behind it. Nscale filed for a US IPO last month naming two anchor customers, and one just walked over a grid connection scheduled for the next decade. Treat interconnect queues, not GPU allocations, as the schedule.
Anthropic offers Blackstone more than 10% to finance $9 billion of chips
Anthropic is offering Blackstone a yield above 10% on its latest chip deal, with Blackstone funds investing $9 billion in debt backed by Anthropic's five-year lease payments on those chips. Source: Reuters This is not an equity round but a lease securitization: a frontier lab paying infrastructure-grade coupons to private credit so the GPUs stay off its own balance sheet.
Judgement: once the cost of a chip depends on your credit spread, model quality stops being the only determinant of unit economics. A five-year lease is a five-year revenue promise, and Anthropic's IPO valuation expectations were trimmed this month. Compare Lambda's $1 billion bet on renting out GPUs: the same compute, financed through very different balance sheets.
CoreWeave's June bonds now yield 15%
CoreWeave's bonds issued in June at 9.675% have plunged, pushing their yield to 15%. Source: Reuters The paper traded below 80 cents on the dollar, the level that marks a bond as distressed, and the shares fell 15.1% from Monday's open through Thursday's close.
The mechanism is the story: rising borrowing costs hit precisely the companies whose expansion is debt-funded. One fund manager told Reuters this is "going to be a turning point in the pricing and availability and structure of credit that's supporting the AI boom". The chief executive blamed a market acting irrationally, which is what executives say when the market is repricing them.
Judgement: this is the AI trade's first margin call, and it landed on the layer assumed safest, the infrastructure everyone expected to be rented out at a spread. If new neocloud debt costs 15% while older paper pays 9.675%, the cost advantage a startup cloud held over hyperscalers is being eaten by refinancing risk.
Startup Snapshot
| Startup | What They Do | Latest Money Event | Valuation / Tape | Why It Disrupts |
|---|---|---|---|---|
| Nscale | Nvidia-backed AI cloud | Microsoft terminated its Loughton contract (Oct 18) | IPO filed Sep 18; $103B contracted value | Private clouds compete on grid connections now |
| Anthropic | Frontier model lab | $9B chip lease debt, Blackstone funds, 10%+ yield | IPO pending; expectations trimmed | Buys compute off balance sheet |
| CoreWeave | GPU cloud for training and inference | June bonds at 9.675% now yield 15% | Public; shares down about 20% | Debt-funded capacity reprices first |
Viral AI: closed doors, metered crawls and a retracted clause
The web is shutting agents out, one account at a time
Users of Meta's Muse, xAI's Grok, OpenAI's Dots and Instinct report blocks when their agents act on sites that used to work. Source: The Verge Yelp admits only agents that pay for its data licensing program, Delta blocks ticket bookings, and JPMorgan Chase and Charles Schwab confirmed they block agents from customer accounts.
Explainer: Muse is Meta's consumer assistant, Dots is OpenAI's agent product and Grok is xAI's. All are sold on the promise that they act for you across the web, which only holds if the other side agrees. Cloudflare spent two years making that optional, and Amazon and Perplexity are still litigating the same boundary.
Meta stopped fighting retail and signed Walmart, Gap, Sephora and Best Buy, while Shopify opened its catalogue to third-party agents. Where access is negotiated, agents work; where it is assumed, they get blocked.
Judgement: Chris Canal of the safety evaluator EquiStamp told The Verge these blocks damage the pitch, because "one of the claims is that they're going to permeate every aspect of our lives". Plan for per-counterparty integrations rather than a universal browser handoff, especially after the rogue agents that read live records.
Meta buried a scraping clause, then quietly removed it
Meta's updated commercial terms, effective October 30, authorized automated software including "web scrapers, bots, robots, spiders, crawlers, and user agents" plus "test credentials" to crawl domains "associated with" a business account and to use that content to "train" Meta products. Source: Nieman Lab Three legal experts told Nieman Lab the language lets Meta argue it can scrape a publisher's site for AI training without a licensing deal, and Cornell's James Grimmelmann said "test credentials" could mean signing up as a fake user.
Meta said the text only consolidated provisions scattered across other terms, that crawling would not change, and that robots.txt still governs access. Then the clause was removed. Source: TechCrunch Ars Technica quoted a spokesperson: "there is no change to how Meta uses web crawlers; any suggestion to the contrary is false".
Judgement: the concession is the tell. A terms-of-service diff is not a negotiation publishers can run at scale. What survives is a rule: robots.txt is a signal, not a wall, so any training pipeline built on crawlable text sits one contract revision away from a licensing invoice.
DeepSeek starts selling search and fetch at 75 cents per 1,000 requests
DeepSeek launched Search and Fetch APIs, giving agents and RAG pipelines live web access from its own index. Source: DeepSeek Search returns ranked results with a 0 to 3 rerank score and an authority tier of 0 to 4, and Fetch returns a page as Markdown. Both bill per successful request at $0.75 per 1,000, whatever the result count, capped at five queries per second, and calls made through DeepSeek Harness, Claude Code and Codex are free until November 1.
Explainer: RAG, retrieval-augmented generation, is the pattern where a model answers from documents it retrieves rather than from what sits in its weights. Search plus fetch as a metered pair is the cheapest defensible version of that pipeline.
The Hacker News thread that carried this to the front page, past 380 points, spent its energy on the arithmetic builders will run anyway: $0.75 per 1,000 calls undercuts your own crawler, proxy pool and extraction stack, provided you accept an index you do not control. It is the same lever DeepSeek's earlier price shock pulled on inference.
Judgement: for regulated buyers, routing the crawl layer through a Chinese API is a compliance decision as much as a cost one. For everyone else, DeepSeek's number is a ceiling on what Western search APIs can charge for the same job.
Viral Snapshot
| Story | What Went Viral | Engagement Signal | What's Actually True |
|---|---|---|---|
| Banks and airlines block agents | Agents failing at checkout and login | The Verge report at 75 comments | JPMorgan and Schwab confirmed the blocks |
| Meta's scraping clause | A terms update read as a land grab | Hacker News 121 points | Meta removed it and denies changing crawls |
| DeepSeek search and fetch | Web access at $0.75 per 1,000 calls | Hacker News front page, 380+ points | Rates published per successful request |
What this means for builders
The two inputs that decide whether an AI product works got repriced this week, and neither of them appears in a benchmark. Credit got more expensive: CoreWeave's bonds yield 15%, Anthropic pays over 10% to lease chips, and Microsoft proved a signed neocloud contract can be walked away from when the grid slips past 2030. Permission got more expensive too.
The consequences are asymmetrical. If you rent capacity, your counterparty's financing now matters as much as its price per GPU-hour, and multi-year commitments are only cheap if the provider can refinance, so shorter terms with renegotiation points are worth the premium. If you build agents, a universal assistant that navigates any site is dead on arrival in the categories that matter most, which moves value to whoever holds the integrations, licensed data and credential management.
DeepSeek's $0.75 per 1,000 calls keeps this from being a scarcity story, because access can be bought at a published rate instead of negotiated site by site. The strategy that survives both trends is unglamorous: use metered APIs for the open web, sign partners for the closed parts, and keep a licensed fallback for the day your crawler appears in someone's terms of service. Whether the credit move is a repricing or a rupture is unresolved, but assume compute gets repriced again before your contract ends.
Related reading
- Nvidia investing in data-center power was the August version of this problem: chips are easy to buy, electricity is not.
- Lambda's $1 billion GPU rental bet is CoreWeave's model on a friendlier balance sheet.
- The rogue agents that read live records is the closest precedent for agents acting without a counterparty agreement.
Frequently Asked Questions
What are the top AI startups to watch today?
Nscale, Anthropic and CoreWeave are the three AI startups to watch on October 18, 2026. Microsoft exited Nscale's Loughton data center after grid officials warned of a 2030s connection, Anthropic is offering Blackstone more than 10% for $9 billion in chip lease debt, and CoreWeave's June bonds now yield 15%.
Why did Microsoft pull out of the Nscale data center deal?
Microsoft terminated its Loughton, Essex agreement because the grid could not deliver power on schedule. The Telegraph reported that national grid officials told Nscale a supply was not expected until the 2030s after repeated project delays. Nscale says its Microsoft relationship remains long-term and that it is exploring alternative power.
Which AI agent story went viral this weekend?
The web's closed doors are the viral agent story: banks and airlines are blocking agents. JPMorgan Chase and Charles Schwab confirmed blocking agents from customer accounts, Yelp admits only licensed agents, and Delta stops ticket bookings. The Verge's report drew 75 comments, and Meta separately removed a scraping clause.
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch ->
