Weekly Digest 39

Topics we are tracking

Amazon Blocks Meta's Muse Agent

Source: GeekWire — Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping

Amazon has blocked Meta's Muse agent from accessing its shopping site, extending the data wars to agents that browse and buy on users' behalf. The two companies remain partners elsewhere: Amazon products have been purchasable inside Facebook and Instagram since 2023.

This is not Amazon's first attempt to control the online shopping experience and customer relationship. Over the past year, it has sued Perplexity over its Comet browser and blocked agents from OpenAI and Google.

Amazon still earns retail margins, marketplace fees and fulfillment revenue when an agent makes a purchase. But it also generated $68 billion in advertising revenue last year. An agent that arrives with a product already selected may bypass the searches and sponsored listings that help generate that revenue.

Amazon has framed the dispute around security and merchant consent, arguing that third-party services making purchases on customers' behalf typically operate with the merchant's agreement. It cited food delivery apps and restaurants as an example. "Agentic third-party applications such as Muse have the same obligations, and we've requested that Meta remove Amazon from the experience," the company said.

Other companies have quickly announced custom connectors for Muse. For these companies, connecting to a personal AI agent can increase the value of their products. If Muse can access my Peloton app and find a class that fits my schedule, that is good for both Peloton and me. I still need a Peloton bike and subscription, so neither revenue stream is disrupted.

The same dilemma is already playing out in publishing, where the discussion is further along. Companies have an incentive to restrict agents when the revenue they lose outweighs the additional business agents bring. But blocking access could also send customers to competitors that welcome their agents. It is not yet clear which companies can afford that trade-off. Nor is it clear how well they can enforce a block as computer-use agents improve.

Muse also needs a lot of computing power. Freda Duan, who writes the Robonomics newsletter, estimates that serving 100 million daily users would take about 1 GW of power in her base case. The virtual computers each agent runs on need only about 0.1 GW of that. Most of it goes to model inference, which grows quickly as agents make more model calls per user.

A stacked bar chart of estimated average power for Muse at 100 million daily users. The sandbox layer of CPUs and memory needs about 0.1 GW at every usage level. Inference at 5 Wh per model call needs 0.5 GW at 25 calls per user per day, 1.0 GW at 50, 2.1 GW at 100 and 4.2 GW at 200; at 10 Wh per call these double. Meta is rumoured to add 7 to 10 GW of compute next year.

Sandbox VMs (CPU + DRAM)Inference at 5 Wh per callExtra if 10 Wh per call

Sandbox at Duan's base case (0.02 to 0.3 GW across low to high cases). Inference at 5 Wh per heavy model call, plus the extra if each call takes 10 Wh. Source: Freda Duan (Robonomics); Diffusion Capital recalculation.

Source: Robonomics · 2026-09-27

SB Energy Postpones Its IPO

Source: Reuters — Data center IPO hopefuls brave tougher market as investor scrutiny grows

SoftBank-backed SB Energy has postponed plans to formally market its IPO, according to Reuters. The company is addressing additional SEC questions while investors scrutinize its proposed valuation and dependence on OpenAI. Reuters reported that valuation discussions had centered on roughly $60 billion.

SB Energy started as a renewable energy play for SoftBank, focusing on solar and battery technology, but has since pivoted into AI data centers. However, in its S-1 published earlier this month, the company disclosed that it has no operating data center today. It stated $439 billion in contracted future revenue, but $357 billion of that will only be recognized eight years from now. The first 50 MW are expected to come online in Q4 2026 at its data center in Cosmos, Texas, with additional capacity coming online in Milam County, Texas, in 2028.

A stacked column chart of SB Energy's contracted data-center capacity by the year it is estimated to reach ready-for-service. Cosmos adds 50 MW in 2026. Milam County adds 154 MW in 2027 and 599 MW in 2028. The PORTS-Pike campus in Ohio adds an estimated 500 MW in 2028, 2,000 MW in 2029, 2,000 MW in 2030, 3,000 MW in 2031 and 524 MW in 2032. The total is 8,827 MW, of which 8,024 MW is at PORTS-Pike.

Cosmos, TX (SoftBank)Milam County, TX (OpenAI)PORTS-Pike, OH (OpenAI)

Site totals, first-phase and final dates are from the S-1. The PORTS-Pike split by year is our estimate, fitted to the backlog the S-1 expects to recognize in each 24-month window. Milam Building 1 is split evenly across 2027 and 2028. Source: SB Energy S-1/A (21 Sep 2026).

Source: SB Energy, Inc. (SEC filing) · 2026-09-21

The S-1 also raises serious questions about SB Energy's proposed valuation. In the first half of 2026, the company reported $138.7 million in revenue and a net loss of $3.21 billion. At a reported $60 billion valuation, that amounts to roughly 216 times annualized first-half revenue. Even that revenue figure included $67.4 million in unrealized gains on power-price derivatives. The loss also requires qualification: $2.57 billion came from warrant revaluation and another $589.5 million from noncash stock compensation.

With current revenue coming largely from power operations, the valuation rests mostly on the future data-center business. Investors are being asked to pay heavily for that income before the company has brought a single data center into operation. Those facilities still need to be financed, completed and powered, and their customers must remain able to pay. The returns left for shareholders after those costs are far from assured.

Political opposition adds another layer of uncertainty. Data centers have become a contentious issue ahead of the midterm elections. Resistance is growing among voters in both parties, even as the Trump administration continues to support their development. Projects that take years to complete remain exposed to changes in permitting rules, power policy and political support through the 2028 election and beyond.

Nvidia has invested in SB Energy and provided guarantees capped at $105 billion for the Ohio campus. But those guarantees take effect in stages, subject to conditions including facilities becoming ready for service. Investors therefore still face the risk of getting the projects built and operational. Nvidia's backing helps the financing and also supports future demand for its chips. Even with that backing, investors have reportedly pushed back on SB Energy's proposed valuation.

Monetary Policy and Long-Term Interest Rates

Source: VoxEU — Anatomy of a rise: Monetary policy and the post-Covid surge in long-term interest rates

US interest rates remain well above their pre-pandemic levels, but there is little agreement on what explains the shift. Inflation and tighter monetary policy account for part of the increase. The harder question is why investors now expect borrowing costs to remain elevated for so long.

Government debt has become a popular explanation for higher yields. Yet its broad trajectory has been known for years, and crossing the $40 trillion mark seems to have attracted disproportionate attention. A debt stock of $39 trillion presents much the same fiscal challenge as $40 trillion; passing a round-number threshold does not suddenly transform the outlook. Fiscal risks can certainly affect bond yields, but explaining a rapid repricing requires identifying what changed in investors' expectations or their willingness to bear those risks. The need to improve public finances does not, by itself, explain why yields have risen so sharply.

A new VoxEU column by Paul Beaudry, Paolo Cavallino and Tim Willems examines whether monetary-policy expectations can keep long-term real rates disconnected from their fundamental drivers. The authors find that three-day windows around US employment reports and speeches by prominent Fed policymakers account for 90.5% of the cumulative increase in ten-year nominal Treasury yields between August 2020 and September 2026. Those windows cover just 23.9% of trading days. A similar concentration appears after adjusting yields for the estimated term premium.

Three weekly lines from January 2020 to September 2026. The 10-year Treasury yield falls from about 1.9 per cent to a low of 0.55 per cent in mid-2020, then climbs to about 5.2 per cent. The 5y5y forward rate follows a similar path and ends near 5.2 per cent. The model-based expected short rate rises less, to about 4 per cent, so part of the rise is term premium.

10-year Treasury yield5y5y forward rateExpected short rate, 10-year average

Last reading of each week. The dashed line marks August 2020, the start of the column's sample. The 5y5y forward is derived from the Fed's zero-coupon curve. The expected short rate is the Kim-Wright model yield less its term premium, so it differs from the column's model.

Source: FRED - 10-year Treasury constant maturity rate (DGS10), Federal Reserve Board - Gürkaynak, Sack and Wright zero-coupon yield curve, FRED - Kim-Wright 10-year fitted yield and term premium (THREEFY10, THREEFYTP10) · 2026-09-25

The authors argue that long-term real rates could remain above or below r* for prolonged periods. r* is the neutral real interest rate, consistent with stable inflation and activity at potential. Expectations about monetary policy could influence those rates without a corresponding change in the underlying fundamentals. A higher long-term real yield would therefore not necessarily mean that the economy's neutral rate had risen. Nonfarm payroll releases and Fed speeches are generally more informative about the near-term path of monetary policy than about changes in the neutral rate.

When higher rates are expected to persist, they reduce asset values, which discourages spending. But households saving for retirement may also need to save less today if their investments are expected to earn higher returns. That leaves more income available for consumption. In the model, these effects can largely offset each other. Rates can then remain away from neutral without a pronounced response in activity or inflation.

This creates a possible feedback loop where the Fed raises rates, the economy remains resilient, and policymakers conclude that the neutral rate must have risen. That interpretation supports keeping rates higher for longer, which in turn reinforces market expectations.

Seen on X

Other interesting stories

• Anthropic: Claude Discovers a Novel Enzyme System
• Google: Behind Project Suncatcher, Our Moonshot to Put AI in Space
• Financial Times: Corporate America Embraces Cheaper 'Open' AI Models
• New York Times: The Surprising Reasons China Is Skeptical of A.I. Safety Calls
• Forecasting Research Institute: How Accurate Have AI Progress Forecasts Been So Far?