Weekly Digest 42

Topics we are tracking

CoreWeave triggers a selloff in the AI trade

Source: Reuters

This was the S&P 500’s worst week since late June. Concerns about CoreWeave’s creditworthiness added to a wider selloff across neoclouds and the AI trade. Most neoclouds lost around 25% this week adding to earlier losses this quarter. CoreWeave ended Friday approximately 67% below its June 2025 high with other neoclouds in a similar range.

Cipher Digital: 62.6%; Applied Digital: 62.3%; IREN: 58.7%; TeraWulf: 54.5%; Core Scientific: 52.8%; CoreWeave: 50.8%; Hut 8: 48.8%; Nebius: 37.4%. Percentage share-price losses over the same June 1–October 16, 2026 period.

Percentage losses in USD closing prices from June 1, CoreWeave’s highest June 2026 close, to October 16. Same starting date for all eight stocks; dividends excluded. Includes GPU-cloud providers and AI data-center hosts.

Source: Financial Modeling Prep · 2026-10-16

Wednesday’s stronger than expected inflation data briefly raised expectations for an October rate hike before those expectations subsequently retreated. Higher interest rates would put additional pressure on an AI trade whose valuations depend heavily on future cash flows. We still consider an October increase unlikely given the midterms are 6 days after the FOMC meeting, but even the temporary repricing was enough to bring longstanding concerns about financing costs and investment returns back into focus.

CoreWeave is both the largest public neocloud but also the company with substantially more leverage on their balance sheet than peers such as Nebius and IREN. In the first half of 2026, it generated $3.7 billion of operating cash flow against $14.1 billion of cash investment in property and equipment. The continued expansion leaves the company dependent on additional financing as the macro environment is increasingly unfavourable and investors are becoming more worried about creditworthiness.

Selected unsecured bonds traded at market yields of roughly 15–16%. That does not mean CoreWeave pays those rates across its existing debt. In its September presentation, the company calculated a weighted-average effective interest rate of 8.3% on its June-end borrowings. Existing fixed coupons do not reset when bond prices fall, although higher yields raise the hurdle for new borrowing and refinancing. Some project loans also carry floating rates above the Secured Overnight Financing Rate (SOFR), exposing unhedged borrowing to higher benchmark rates.

The company reported approximately $104 billion of backlog at June-end, which far exceeds the backlog of other neoclouds. These contracts provide visibility, but collecting the revenue requires delivering capacity and paying operating costs. CoreWeave also consistently maintained the highest ranking in the ClusterMax Benchmark from SemiAnalysis showcasing their expertise in building and managing clusters. Customer concentration is another worry that investors have raised which is making CoreWeave dependent on a relatively small number of buyers. However a successful OpenAI IPO (OpenAI is the largest CoreWeave client) and continued strength among its largest customers would support its financing case. The current narrative is swinging more in favour of open-weight models rather than frontier models which could put margin pressure on OpenAI and Anthropic but also increase the demand for independently hosted inference.

In our view, an imminent CoreWeave bankruptcy is unlikely, and we expect strong revenue growth in its next results. The more immediate test comes next week, when both SpaceX and Meta report earnings. Their capital-spending plans and commentary on AI returns will help investors assess whether demand commitments remain intact.

Anthropic study on productivity gains and the human bottleneck

Source: Anthropic

Anthropic published an interesting study this week looking at how AI is changing work inside the company. Anthropic is likely at the forefront of internal AI adoption and could provide valuable insights into how a company is built around artificial intelligence. The findings suggest substantial gains in output across both technical and non-technical roles. Code commits per engineer rose roughly 5x over six months, while lawyers processed more contracts and salespeople closed more deals.

One of the most interesting stats for us was that, by August, employee-directed agents were logging six hours of runtime for every estimated hour of human work. This points toward scaling AI through coordinated swarms of agents, which we believe will be a leading theme going forward. We plan to write a dedicated note on this in the coming weeks. Employees at Anthropic describe their work shifting from doing individual tasks to directing and supervising agents. More time goes into setting instructions, reviewing results and deciding which problems require attention. In interactive sessions, agents reportedly spend around 80% of elapsed time waiting for human input. Human attention and judgment are already bottlenecks in our own work, and it will be interesting to see how this changes as we scale agents further.

Approximate digitization of Anthropic Figure 2 on a logarithmic axis. Research organizations reach a reported 19 agent hours per estimated human hour, and all employees reach 6. Displayed dates and curves are estimates, not original observations.

Research organizationsAll employees

Curves and dates approximately reconstructed from Anthropic Figure 2; endpoint labels as reported. Interactive sessions with pooled subagents. Runtime includes inference and tools, excludes waits >30 min; human hours assume 40 h/week per employee.

Source: Anthropic · 2026-10-16

The report also highlights changes in hiring and the skills companies look for. Anthropic reports that managers increasingly emphasize autonomy and judgment over narrow technical expertise and execution. The ability to define a problem and recognize a good result could become more valuable as agents take on more of the implementation. For now, technical knowledge is still needed to assess agents’ work, even if only directionally. But if AI reaches levels where only the top 0.01% of experts in each field can verify its work, we could increasingly find ourselves unable to verify or even comprehend what it produces.

Anthropic’s study does not establish how quickly these changes will spread across the economy. We are already seeing political and intellectual backlash, including from domain experts whose work has changed dramatically and will continue to change. If AI-driven unemployment rises, we expect that backlash to intensify rapidly.

Inflation and the next Fed decision

Source: Bureau of Labor Statistics

The Wednesday inflation report showed US consumer prices rising 0.6% in September, bringing annual inflation to 3.7%, up from 3.4% the month prior. Core inflation came in at 0.3% against the 0.2% expectation, bringing the annual core number to 2.6%. The monthly increase was unchanged from August, but core inflation coming in above expectations was perhaps the bigger surprise than the headline number itself. Core inflation excludes food and energy.

Twenty-six selected CPI categories ranked by year-over-year price change. Year-over-year changes for September 2026 range from fuel oil at 74.2% to motor vehicle insurance at -5.0%. The headline all-items index is 3.7%. These are price changes, not contributions.

Category selection follows Charles Schwab’s CPI chart. Unadjusted CPI-U price changes in September 2026 from a year earlier. Categories overlap and should not be summed as inflation contributions.

Source: U.S. Bureau of Labor Statistics · 2026-10-14

The breakdown of headline inflation clearly showed some of the main culprits behind higher prices. On a year-over-year comparison, fuel oil is up 74.2%, gasoline is up 34.2% and airline fares are up 22.4%. These three accounted for roughly 1.35 percentage points, or 37%, of the overall annual increase in headline CPI. All three are exposed to higher oil prices following the escalation in the Middle East. Rate hikes cannot directly address these supply disruptions, which raises questions about their effectiveness against the current drivers of inflation. Further hikes would also increase borrowing costs for households and businesses already facing higher energy costs. However, interest payments are excluded from CPI, so higher car-loan, mortgage or business-loan payments do not automatically create a further increase in measured inflation. Higher rates can still bring inflation down by reducing demand.

The increase in annual core inflation does complicate the picture because it raises concerns about a passthrough effect into the wider economy. Economists have pointed to three factors impacting the inflationary environment: oil, immigration and tariffs. In our view, all three are influenced by policy decisions coming from the White House, and at least some of these inflationary pressures could be avoided or eased.

Kevin Warsh and the Fed face these dynamics with limited ability to address their underlying causes, but nevertheless have a mandate to maintain price stability alongside maximum employment. An October hike still seems very unlikely to us. Hiking rates six days before a midterm election could be perceived as a political act, although the Fed has acted this close to an election before. This leaves us with the December meeting, where we believe a hike is likely based on the information currently available. Most participants in the September meeting already considered another increase likely appropriate before year-end.

However, there is still one more CPI release, two PCE inflation reports and two monthly jobs reports before the December decision, as well as the aftermath of the midterm elections. We expect a Democratic House, with a Democratic Senate also possible. This could lead to a scramble by the White House and congressional Republicans to rush through legislation before losing their majority. Republicans are already considering raising the debt ceiling during this period to prevent Democrats from gaining leverage over a debt-limit fight expected in 2027. Passing it could prove contentious, and any resulting disruption in financial markets would also matter for the Fed.

Seen on X

Other interesting stories

• Wall Street Journal: Michael Dell, a Dot-Com Era Survivor, Has Tips to Beat an AI Bust
• New York Fed: Has the AI Boom Impacted the U.S. Current Account?
• Justin Mc: Anduril and the Limits of Build First, Buy Later
• CNBC: AI companion startup Character.ai files for Chapter 11 bankruptcy, citing costs and litigation
• ChinaTalk: Why a Chinese AI Lab’s Product Lead Thinks AGI is “Inevitable”
• SemiAnalysis: ASML Is Building a Hybrid Bonding Tool, Sending More Chills to Besi