Insurance on AI debt costs the most since 2021, Aramco takes half of HUMAIN, and 3 more AI moves
Default insurance on Amazon, Alphabet, Microsoft, Meta and Oracle hits its highest since 2021 as a compute-derivatives market forms, Aramco takes 49% of HUMAIN for $2.6 billion, Meta quietly builds a humanoid with Fourier, DeepSeek open-sources a self-starting coding agent, and chatbots flunk an election test two days before the US midterms.
This issue covers October 31 at 12:00 UTC through November 1 at 13:00 UTC — 25 hours, because US Eastern time fell back an hour overnight.
Five developments cleared the bar in that window.
The cost of insuring against a default by Amazon, Alphabet, Microsoft, Meta or Oracle has risen to its highest level since 2021, and a market for hedging the price of computing is forming behind it.
Aramco took 49% of Saudi Arabia's state AI champion for $2.6 billion in cash and assets.
Meta built a working humanoid robot with a Chinese partner in about three months, and is already on a second one.
DeepSeek published an open-source coding agent that decides on its own when a developer's message is a request for work.
And two days before the US midterms, a survey of eight chatbots found almost 40% of their election answers contained at least one factual error.
In brief
- Markets: The cost of protecting against debt defaults at Amazon, Alphabet, Microsoft, Meta and Oracle climbed over the past month to its highest level since 2021, the Wall Street Journal reported on Saturday. Oracle's default insurance widened by more than a percentage point, and trading firms have begun quoting derivatives on the price of renting Nvidia chips. 12
- Funding: Aramco closed the purchase of a 49% stake in HUMAIN, the AI company launched by Saudi Arabia's sovereign wealth fund in May 2025, for about 9.9 billion riyals ($2.6 billion) in cash during 2026 plus AI and computing assets. The Public Investment Fund keeps 51%. 3
- Safety: With the US midterms on Tuesday, an evaluation firm found almost 40% of chatbot answers to election questions contained at least one factual error across eight models, and a disinformation monitor found Google's AI Overview repeating a likely Russian-forged claim about a Senate candidate after fact-checkers had debunked it. 8
Five developments
1. The AI trade now has an insurance market
The Wall Street Journal reported on Saturday that the cost of protecting against a default by Amazon, Alphabet, Microsoft, Meta and Oracle widened over the past month to its highest level since 2021, according to the bond data provider Solve. 1
That cost is a credit-default swap, an insurance contract that pays its buyer if a named borrower fails to repay its debt. The price is quoted as a spread, so a widening spread means the market is charging more to carry the same risk.
Solve co-founder Eugene Grinberg told the Journal that all five had widened together. "There were no exceptions, suggesting broad market concern about the sector rather than idiosyncratic credit events at individual companies," he said. 1
Oracle moved furthest. Its default insurance widened by more than a percentage point over the month. 1
Equity hedging has followed. Trading in put options on CoreWeave, the heavily borrowed AI cloud company, reached its highest level on record in October and stayed elevated, according to Cboe data. The most popular contracts expire in 2028 and pay out if the shares fall to $35; CoreWeave traded around $55 this week. 1
Behind the hedging sits a market in the thing being insured. Jump Trading has opened a desk that trades computing power and structures over-the-counter contracts against it, including protection against a collapse in prices. 1
"We're seeing really significant demand from folks looking to protect downside," said Stephen Solaka, a senior vice president at Gateway Investment Advisers, which runs portfolios for individual clients with concentrated tech holdings. Some clients, he said, have quit their jobs to hedge their own stock. 1
The reference prices come from chip rentals. The cost of renting an Nvidia B300 chip climbed about 56% this year to $7.56 an hour, according to an index from Silicon Data.
A neocloud renting out those chips could buy a put that pays if hourly rates fall to, say, $4. 1
Susquehanna International Group has had enquiries from data centres and private neoclouds, both directly and through brokers, about hedging chip exposure. Nick Volbi, who runs institutional prediction markets at the firm, said a put tied to a compute index can also serve as collateral.
"I'm going to go to a potential lender and say, 'Hey, if I buy this structure and I get this form of insurance, can you now give me financing?'" he said. 1
Even politics is tradeable. Kalshi offers contracts on which states might halt new data-centre development before 2028, including New Hampshire, Michigan and Maine, though trading has been thin. 1
The context is a bad month. The Nasdaq fell 2.9% on Friday and closed its worst month since March 2025. 2
The hedging remains small beside the bet it covers. Jump president Dave Olsen put it this way: "I don't think my house is going to catch on fire, but I still have insurance." 1
2. Aramco buys half of Saudi Arabia's AI champion
Aramco completed the purchase of a 49% stake in HUMAIN through a wholly owned subsidiary, the company said on Sunday, with the Public Investment Fund keeping 51%. 3
The consideration is about 9.9 billion riyals, or $2.6 billion, in cash contributions during 2026, plus AI, computing and digital assets transferred into the company. 3
The deal formalises a non-binding term sheet the two signed on October 28, 2025, under which their AI assets and talent would be combined under HUMAIN. 9
HUMAIN launched in May 2025 and is building data centres, cloud capacity, models and industry applications. Its flagship model is ALLaM, a multimodal Arabic language model. It targets three gigawatts of AI data-centre capacity by 2030. 3
Aramco's chief executive, Amin Nasser, framed the stake as demand as much as investment. He said the company had booked $11.3 billion of "technology realized value" since 2023, with a significant portion from AI, and that it expects to run industrial applications on HUMAIN's infrastructure. 3
The fund's side is consolidation. Yazeed Al-Humied, who heads Middle East and North Africa investments at PIF, said bringing the two portfolios together under HUMAIN creates "a powerful synergy that fuels talent growth, technological advancement, and intellectual property creation". 3
Capacity is already being sold. On October 21, HUMAIN and DataVolt announced a preliminary framework with the AI cloud company Nebius for up to 50 megawatts at a planned campus in Neom's Oxagon.
That is half of the 100 megawatts under construction in a first phase planned for 360 megawatts, with capacity due in stages from the fourth quarter of 2027. 9
For readers tracking who funds AI compute outside the United States, the closing changes the ownership map. Two Saudi state entities now hold one company between them: an energy producer with an industrial customer base, and a sovereign fund with a mandate to build national champions.
3. Meta has a humanoid, built in three months
The Information reported on Saturday that Meta built a humanoid prototype with Fourier, a Shanghai robotics company, and has begun a second version. Reuters relayed the report and said it could not immediately verify it. 45
The prototype is called Max inside Meta. It weighs about 106 pounds (48 kilograms), stands roughly 1.5 metres tall, and runs on Nvidia's Jetson Thor chip. 10
Its head is a screen that mimics facial expressions, and it carries speakers and five-fingered hands built to pick up small objects. 10
The build took about three months. Meta's vice president of product for robotics, Li-Chen Miller, spent almost two months at Fourier's factory in Shanghai while the two teams worked, and Meta showed Max to some employees at an internal robotics summit in late August. 510
Work on Max 2 began in September. The second unit is meant to get a redesigned torso, better hands, a projector and extra sensors for stairs. 10
Meta's stated purpose is research. The company told The Information the work with Fourier is for internal experimentation and that it has no plans to sell the units, and one of the report's sources said there is no launch date for a humanoid product. 10
The strategic logic points at Meta's assistant. The company describes humanoids as a possible physical extension of its Muse AI agent, aimed eventually at household tasks, and it bought the San Diego robot-software startup Assured Robot Intelligence in May to staff the effort. 10
Fourier has been building humanoids for years. It unveiled GR-3 in August 2025, a 165-centimetre machine it markets as a care companion, and it has been selling into that market while larger manufacturers chase factory work. 10
What the report leaves open is the part that decides whether this becomes a product: cost per unit, whether Meta intends to manufacture, and which tasks Max completes without a person guiding it.
For now the confirmed claim is narrower — a large AI company can assemble a humanoid in a quarter, with a partner, and keep it in the lab.
4. DeepSeek open-sourced an agent that starts work on its own
DeepSeek's Harness team published a design essay, "Exploring Super-agent Design: Towards a More Autonomous Agent System", and released the corresponding code on Saturday as dsh-superagent-prototype, under the MIT licence. 67
The repository went public on Saturday and passed 5,500 stars and 500 forks within a day. 6
What it does: it watches one GitHub repository, and optionally a Feishu group chat. When a person the team has approved asks for work, the agent opens a ticket.
The ticket runs in its own session and its own directory, and the answer comes back in the original conversation, usually as a pull request. 6
The architecture splits two jobs. A long-running router handles conversation and dispatch; separate worker sessions do the work, so one slow task cannot block the next request. 11
DeepSeek's stated design goal is human attention. The team argues that a person can follow only three to five agent sessions at once, which makes operator attention the binding constraint on agent output. 11
Proactivity is gated by three modes — off, observe and auto — and observe is the default, meaning the agent records the work it would have accepted without doing it. 11
DeepSeek's own performance figures come from a simulation: streams of about 300 messages per run, roughly 20 of them real requests, five runs per setup. The agent finished tasks in a 95th-percentile time of 15.7 minutes against 29.5 minutes for a single-session baseline, with a median of 12.4 minutes. 11
Those numbers measure scheduling against synthetic traffic. DeepSeek has yet to publish a benchmark of the patches it writes. 11
The repository says as much, at the top of its README. DeepSeek calls it "research prototype for reference only, not an official DeepSeek product", and warns that much of the engineering was done by coding agents.
The published code also differs from the internal version described in the essay. It runs with the full permissions of the operating-system user who starts it, so anyone trying it should isolate it first. 611
Bloomberg framed the release as a challenge to Anthropic's Claude Code, and the comparison extends to the licensing. Claude Code is a paid product sold with a subscription; this is source code anyone can run, pointed at whichever model their harness profile selects. 7
5. Two days before the midterms, the chatbots get graded
Voters go to the polls on Tuesday. AFP reported on Sunday that this is the first federal election in which millions of voters are actively relying on AI tools for political information, and on what the research so far shows. 8
Forum AI, a startup that evaluates models, tested eight of them on election questions and found that almost 40% of the answers contained at least one factual error. 8
The errors were specific. Asked why voters should elect Representative Don Davis of North Carolina, Anthropic's Claude listed "former school superintendent" among his credentials, describing someone else.
Asked about voting in Pennsylvania, Google's Gemini 3.6 Flash gave November 3 as the deadline to request a mail-in ballot. The real deadline had passed a week earlier. 8
"A voter who followed the answer would have asked for a ballot too late to use it," the study said. 8
The models were worse on logistics than on candidates, which is the reverse of what one might expect, and they put "the wrong words in the mouths of candidates, officials, and observers hundreds of times", often sourcing claims to low-quality pages such as Reddit threads. 8
Anthropic did not respond to AFP's request for comment. Google said it was focused on helping voters find reliable information and pointed to quality and policy protections across its AI products on election topics. 8
Foreign operations have already worked their way in. In late October, AFP's fact-checkers debunked a likely AI-altered clip that appeared to show a former student of Texas Senate candidate James Talarico accusing him of sexual misconduct, an allegation that carried the marks of a Russian campaign.
NewsGuard found Google's AI Overview repeating the claim, as did Gemini and Copilot. 8
Both companies said the systems updated. Google said the issue was addressed as its systems surfaced the latest reporting, and Microsoft said Copilot reflected the updated context and identified the claim as fabricated once more information was available.
A later AFP test found the tools citing professional fact-checkers, which shows how much the answer depends on when the question is asked. 8
Use is broad and trust is thin. In a National Conference on Citizenship survey, two in five Americans said they had used AI tools to look for election information.
Only 40% trusted them with logistical questions such as when, where and how to vote, and 60% trusted them "only a little or not at all" to be unbiased about candidates. 8
There is counter-evidence. The Brennan Center for Justice found that leading chatbots pushed back against common conspiracy theories, including claims that voting machines were rigged, even when users pressed them repeatedly. It said they behaved as sympathetic listeners to sceptical users while refusing to endorse the theories. 8
The practical reading for Tuesday is narrow. For a candidate's record or a ballot deadline, check the state's own page. For a viral claim about a candidate, check whether a fact-checker has published on it, and re-ask the assistant afterwards.
The read-through
The five stories sit on one axis: what the build-out costs to keep running, and who is left holding the risk. That framing is my inference from the five items.
Start with the money. The companies spending most on AI now pay more to insure their debt than at any point since 2021, and the new part is the machinery around that price — compute derivatives, chip-rental puts, hedges that a lender will accept as collateral. 1
A market in insurance usually appears after the risk is visible. Jump's own president frames the demand as buying insurance, and the volumes are small next to the positions they cover. 1
The commitments keep arriving anyway. Aramco put $2.6 billion of cash and assets into a company whose plan requires three gigawatts of data centres by 2030. 3
The product layer moved just as fast this weekend: one large lab took a humanoid from nothing to prototype in a quarter, and a Chinese lab gave away an agent harness that competes with a paid product. 610
The fifth story is where the cost lands first. Models that answer election questions wrongly at a measured rate, days before a vote, are the same models being sold into workplaces and households, and the fix so far is a company response after a fact-checker publishes. 8
Held together, the weekend reads as a sector pricing its own risk while it keeps building. Whether that is prudence or a late start is the question the next few weeks of credit spreads will answer.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.