Weekly Digest 41

Topics we are tracking

New Open Source Model Releases

Sources: Reflection AI — Introducing Beam: Reflection’s 501B open-weight model; Mistral AI — Introducing Mistral Large 4

Reflection AI and Mistral introduced new models this week, pitching them as alternatives for businesses and governments seeking more control over their AI. Reflection’s first model, Beam, targets coding, reasoning, and tool use. Mistral Large 4 adds image understanding and emphasizes professional workloads.

Both models are impressive, but neither establishes a clear overall lead over the strongest Chinese alternatives. Artificial Analysis places Mistral 4 roughly alongside DeepSeek V4.1 Flash on general intelligence, but at a higher cost. Among open models, Mistral is the best non-Chinese model, but we are unsure how much this really matters if five Chinese companies have created better open models. Beam has not yet been scored on Artificial Analysis’s Intelligence Index, but based on the benchmarks published in Reflection’s blog post, we would not expect it to outperform Mistral 4.

Artificial Analysis Intelligence Index: open-weight models

46

45

44

42

39

38

~35

34

31

29

25

23

MiMo-V2.6-ProXiaomi

GLM-5.3Z.ai

Kimi K3Moonshot

GLM-5.3-FlashZ.ai

DeepSeek V4.1 FlashDeepSeek

Mistral Large 4Preview

Reflection Beamour estimate

Qwen3.8 27BAlibaba

K2 HorizonMBZUAI

MiniMax-M3MiniMax

InklingThinking Machines

Nemotron 3 UltraNvidia

In both launch posts, the companies highlight that these models were trained in the West and emphasize the sovereign AI angle. American and European companies can already run Chinese-developed open models on their own or rented infrastructure without sending their data to the model developer. Microsoft, for example, explicitly states that prompts and outputs in the relevant Azure-hosted deployments are not shared with the model provider.

Benchmark scores, cost, and country of origin should not be the only reasons to choose one model over another. Thinking Machines, for example, made that case explicitly with Inkling by linking the model to its fine-tuning service through the Tinker API. We have found this combination very useful internally and think it fills a niche. Mistral has launched something similar with its Forge service and has named ASML, Ericsson, and the European Space Agency among its partners. Both companies are betting that more businesses will adapt models using their proprietary data, achieving better results on specific tasks at a lower cost than frontier models. We don’t think that is an unreasonable bet.

After the Reflection AI launch, the FT also reported that Nvidia (who already invested in Reflection) is discussing deeper ties with the company including a potential acquisition. If an acquisition goes ahead, it would be another model developer that Nvidia acquired after the Poolside AI reverse acqui-hire in August and it would establish Nvidia among the top American Frontier labs.

SpaceX and SoftBank seek more capital for their AI ambitions

Sources: Financial Times — on SpaceX’s chip financing; Financial Times — on Masayoshi Son’s Gulf fund

SpaceX is seeking an additional $40 billion to purchase Nvidia chips through a financing package expected to be led by Apollo Global Management, according to the FT. The proposed structure includes roughly $10 billion in bank loans and $30 billion in investment-grade debt, with Pimco among the potential lenders. This comes after $87 billion raised in the IPO and an additional $25 billion bond raise shortly after.

In its latest reported quarter, SpaceX generated $2.42 billion in operating cash flow, but spent $18.37 billion on capital expenditures, leaving approximately $15.95 billion in negative free cash flow. The company is currently valued at approximately $2.1 trillion, making it the seventh-largest publicly traded company in the world, immediately behind TSMC. TSMC generated approximately $24.8 billion in operating cash flow and $9.1 billion in positive free cash flow during the same quarter.

Using its June balance sheet as a baseline, an additional $40 billion of debt would increase its debt-to-book-equity ratio from approximately 0.30 to 0.62, roughly doubling it, assuming equity and other debt remain unchanged. SpaceX’s five-year credit-default-swap spread rose by as much as 16.7 basis points to 197.6 basis points during trading on October 7, following reports of the proposed financing.

Stacked bars of hyperscaler bond sales by company. In 2025 Alphabet sold $38bn, Meta $30bn, Oracle $26bn and Amazon $15bn, $108bn in total. From January to 11 October 2026 Amazon sold $98bn, Alphabet $81bn and Meta, Oracle and SpaceX $25bn each, $254bn in total. We estimate $453bn in 2027: Alphabet $140bn, Amazon $134bn, Meta $65bn, Oracle $60bn and SpaceX $53bn.

AlphabetAmazonMetaOracleSpaceX

2025 and 2026 (to 11 October) are actual bond sales in all currencies. 2027 uses Goldman Sachs’s $400bn forecast for the five cloud companies, about 35% of their planned capex, split by each company’s past share of sales. SpaceX is our estimate. Microsoft has sold no bonds since 2017. Off-balance-sheet data-centre deals are not included.

Source: Amazon, SEC EDGAR pricing term sheets and debt footnotes for AI-buildout bond issuers, Goldman Sachs, Diffusion Capital · 2026-10-11

We hold no position in SpaceX, either long or short, and are watching this from the sidelines for now. Musk’s track record gives us reason to take these ambitions seriously, but we honestly don’t know whether SpaceX can scale this quickly—or whether tighter financing conditions will catch up with it first. SemiAnalysis argues that SpaceX could reach around 10GW of data-center capacity by the end of 2027, using on-site power generation, alternative equipment suppliers, and faster construction methods to deliver capacity ahead of competitors. In a market where customers urgently need compute, that speed could command a substantial premium. Under its assumptions about delivery, pricing, and the share of capacity rented to customers, SemiAnalysis sees a path to a $300 billion annualized revenue run rate by the end of 2027. That is a very ambitious modeled scenario but it could explain why borrowing heavily could make commercial sense. SpaceX does not necessarily have to win the Model race with Grok but instead could either focus on supplying infrastructure to other AI companies or lean into the Application layer with Grok Bot.

Separately, SoftBank Group founder Masayoshi Son is reportedly aiming to raise up to $100 billion from investors in the Gulf region to build a fund to buy companies and then make them more efficient using AI and similar technologies. The FT reported that his robotics and physical AI venture Roze will also be involved in this venture. This report follows on the heels of SoftBank-backed SB Energy's reported delay in its IPO as investors balked at the $50 billion or so valuation. In a post on X Elon reacted to the story with a comment: "Only a fool would give them money."

Amazon’s data centers face limits on grid power

Source: The Information — How Amazon Rewired Its Data Centers After Mississippi Power Warning

Amazon has adapted its approach of switching its Mississippi data centers to backup power in the wake of warnings from Entergy that it might need to disconnect the facilities during periods of peak demand. Reports suggest Project Falcon allows Entergy to trigger the switch itself, rather than having to require AWS to initiate the process, but the issue is that there isn't enough capacity on the grid at peak.

This situation is not unique to Mississippi: Texas enacted similar laws that allow ERCOT to order large customers to install backup power or reduce demand when there's an emergency. Duke Energy in North Carolina is said to have added curtailment clauses to contracts with large customers, and PJM is pursuing similar arrangements that would curtail large loads during shortages. All of these are different, but all seek to impose conditions on the developer's access to power.

It's expensive to add new supply to the U.S. electrical system, and we see a strong case for investing in future proofing the US grid. Data center power demand could provide funding to make that necessary investment. Electricity consumption in the U.S. remained broadly flat through the mid-2000s to early 2020s. That's made adding new capacity to the grid more urgent. New demand from hyperscalers and data center developers could provide the funding and long-term commitments to expand supply. If these large scale customers are willing to pay the additional costs they create and invest in shared infrastructure, then this could mean lower bills for other customers.

US retail electricity sales by sector, 2000 to 2040. Total sales barely moved between 2005 and 2021, then rose to 4,058 TWh in 2025. On EIA's forecasts, commercial sales, which include most data centres, grow from 1,592 TWh in 2027 to 2,062 TWh in 2040, faster than residential or industrial sales, and total sales reach about 5,095 TWh. If demand grew as fast as utilities forecast to 2030, commercial demand would be about 640 TWh higher in 2030 and about 794 TWh higher in 2040.

ResidentialIndustrial and transportCommercial, incl. most data centresExtra commercial demand if utility forecasts hold

2000–25 actual; 2026–27 EIA’s October Short-Term Energy Outlook; 2028–40 extends each sector at the growth rates in EIA’s 2026 Annual Energy Outlook. EIA does not report data centres separately; most sit in the commercial sector. The top layer is our calculation: it adds the gap between EIA and the 5.7% a year that utilities’ forecasts imply for 2025–30 (Grid Strategies), assigns it to the commercial sector, and grows it with EIA’s commercial path after 2030.

Source: US Energy Information Administration, Grid Strategies · 2026-10-06

Of course, data center reliance on electricity raises the question of backup power, which has its own challenges. Generators require fuel, maintenance, and permits to operate. They're also polluting and may face greater community opposition when used more frequently, even for facilities where self-generation is the main power source. Oracle's Project Jupiter in New Mexico has shown what kind of issues can arise. The gas pipeline intended for the campus has been delayed until February 2027, versus August 2026, due to permitting issues, and the state land commissioner rejected applications for part of the route in March and again in July. Oracle later proposed an alternative plan with Bloom Energy fuel cells, rather than turbines and diesel generators, but still depends on securing gas.

We think Falcon could be a practical response to temporary shortages, but calling it a better economic model would be premature. Grid constraints and the backlash over electricity bills are making on-site generation more attractive, yet producing electricity independently brings its own infrastructure and permitting problems. There is therefore a case for data centers remaining connected, helping finance improvements, and reducing their draw when the grid is under pressure. Whether that becomes a workable compromise depends on how often restrictions occur, what reliable backup costs, and whether the agreements protect other electricity customers from paying for the expansion.

Seen on X

Other interesting stories

• Daron Acemoglu: Will AI Replace Workers? Not If We Build It Right.
• Toby Ord: AI Scaling vs Human Scaling
• New York Times: Global Trade Proves Resilient, WTO Says
• Paul Krugman: What’s the Matter With France?
• Nathan Benaich and Air Street Capital: State of AI Report 2026
• The Substrate: Having the Best Robotics Model Isn’t Enough to Win the Robotics Race