Sunday, November 8, 2026

@StableBread Oct 11 2 clicks

Fajasy on X

AI labs are paying to close the gap on malware binaries their best models still can’t read. A ransomware nanocap I wrote up last month is now training its own model to read them, after data labeling companies approached its team. Even with far more capable models today, OpenAI and other top AI labs still pay experts to write and grade the problems their models get wrong. And that’s where data labeling companies play a role. They recruit experts in fields where models still make mistakes, like coding, finance, and medicine, and turn what those experts know into graded examples for the labs to train on. Here are just some of the valuations of the largest AI data labeling companies: → Scale AI: Valued at over $29B after $META paid $14.3B for a 49% stake in June 2025. → Surge AI: Was in talks to raise at a $25B valuation in July 2025. → Mercor: Was in talks to raise at a $20B valuation in July 2026. One of the skills the labs need help with is reading malware, the malicious software hackers use to break into or lock up computers. Malware often arrives as a binary, a compiled program in machine code, and many are obfuscated, or scrambled to hide their purpose. In August 2026, researchers led by Columbia University released SRE-Bench, a test of AI models on 262 binaries, and concluded that reverse engineering, or figuring out what a binary does, “remains largely unsolved.” According to Vals AI’s SRE-Bench leaderboard, OpenAI’s GPT-6 Astra claims the top spot, fully solving 56.9% of the binaries. But GPT-6 Astra still can’t fully solve the other 43.1%. Data labeling companies are now recruiting malware experts to train the labs’ models. Scale AI, for example, is hiring a product manager to build a cybersecurity product line, with “malware and binary analysis” among the skills it covers. Mercor also pays security experts $200-250/hour to evaluate AI-generated analyses, preferring people with reverse engineering or binary analysis backgrounds. That demand for malware skills has now reached the ransomware nanocap I wrote up last month. Its CEO said in a new interview that data labeling companies have asked the company to help label what each malware binary does, like whether it’s secretly encrypting files. Its software copies the encryption keys ransomware creates during an attack, so its engineers can build a decryptor (a program that reverses the encryption) and recover a victim’s files without paying the ransom. So is a company the market values at under $10M on track to be acquired by, or partner with, an AI lab or data labeling company? Or is binary analysis another AI plan that misses its targets, like the AI decryptor builder it has pitched since 2023? That's what I talk about in my ~3,500-word company update: https://t.co/IGkfgBX5hR

@Fierce__beast Oct 11

Fierce_beast on X

How frequently do restructurings at this scale actually work? Did it work for $STLA? https://t.co/TcKUd8hB3M

@StableBread Oct 11

Fajasy on X

Continue reading here for the U.S. AP1000 push and the new plant $BWXT is weighing: https://t.co/bPIYua98gd

@StableBread Oct 11

Fajasy on X

Excluding medical isotopes, commercial power revenue at BWX Technologies $BWXT nearly tripled, from ~$400M in 2023 to ~$1.1B expected in 2026. $BWXT’s commercial segment is growing much faster than its government segment, driven by (1) Ontario rebuilding its reactors, (2) Canada’s plan for new ones, and (3) the U.S. push for new AP1000 reactors. Its commercial backlog more than doubled to $1.6B. Kinectrics, a Toronto-based nuclear engineering and testing firm acquired in May 2025, explains part of the jump. Commercial revenue grew 72% in Q2 2026, or 33% excluding the acquisition. In Ontario, utilities are rebuilding reactors to run for decades longer. $BWXT is building 48 steam generators for Ontario Power Generation’s (OPG) Pickering refurbishment, part of more than CAD$1B (~$700M) of contracts. The same contracts cover the reactor pressure vessel for the first BWRX-300 at OPG’s Darlington station, a 300 MWe small modular reactor (SMR) designed by GE Vernova Hitachi. The 10-K calls $BWXT “the only commercial heavy nuclear component manufacturer in North America.” Notably, its Cambridge, Ontario plant is turning out more than 1.5 steam generators/month vs. the 1/month its schedule requires, per $BWXT’s Investor Day on September 29, 2026. Beyond refurbishments, Canada’s June 2026 nuclear strategy calls for “up to ten new large-scale reactors,” with two under construction by 2035 and five more planned or under development by 2040. $BWXT already has licensing and early engineering contracts for the large reactors planned by OPG and Bruce Power, the private operator of Ontario’s Bruce nuclear station. John MacQuarrie, president of Commercial Operations, estimated Canada’s new-build market at CAD$120-150B (~$84-105B) at Investor Day. Slide 31 shows $BWXT’s potential revenue per new reactor: → CANDU (Canada’s heavy-water reactor design): $350-700M per reactor. → AP1000 (Westinghouse’s large U.S. design): Low hundreds of millions of dollars. → SMRs: $50-100M.

@DevinLaSarre Oct 11

Devin LaSarre on X

Newest out. Imperial Brands: Collapse $IMBBY $LOG $MO $PM $BTI $STG $HAYPP https://t.co/n7OxkJ7DQj

@convequity Oct 11 6 clicks

Convequity on X

OpenAI and Anthropic will rent most of the AI capacity they plan, but mainly from the big clouds rather than the listed neoclouds Convequity's AI Bubble Barometer (ABBX) tracks the data-centre capacity each company plans and how much of it it expects to rent from partners rather than own. 1. OpenAI plans 21.2 GW, of which 18.0 GW is to be supplied by partners. Anthropic plans 13.8 GW, all of it from partners. Together the two labs plan to rent 31.8 GW. 2. That is 2.8 times the 11.2 GW that the five hyperscalers, $MSFT, $AMZN, $GOOGL, $META and $ORCL, plan to rent from others. 3. The labs' named suppliers are mostly AWS, Oracle, Google, Azure and SB Energy. CoreWeave, a listed neocloud, and Nscale, a private one, appear on the lists, but the bulk of the 31.8 GW is not with listed neoclouds. 4. The seven listed neoclouds, $CRWV, $NBIS, $IREN, $APLD, $CORZ, $CIFR and $BTDR, plan to build 33.62 GW of their own. Much of that is gross site power rather than contracted IT capacity. 5. The labs are also behind most of the recent growth in AI capacity plans. Since 14 Jul, the total capacity that the 15 ABBX companies say they plan to have has risen by 22.19 GW, from 131.23 GW to 153.42 GW on 7 Oct. OpenAI and Anthropic raising their own plans accounts for 13.7 GW of that rise (62%). We think the labs' rental demand mainly supports $AMZN, $ORCL, $GOOGL and $MSFT. The neoclouds will need to fill their pipelines largely with other customers, including the hyperscalers' own 11.2 GW of planned rentals, so we would not read the labs' plans as demand for the neoclouds.

@StableBread Oct 11

Fajasy on X

The Department of Energy conditionally committed $17.5B in loans in June 2026 for long-lead equipment for up to 10 AP1000 reactors (Westinghouse’s large U.S. design), and only “a limited number of heavy engineering plants” can make the components, per World Nuclear News. BWX Technologies $BWXT bought Precision Components Group (PCG) in July 2026 to get a U.S. foothold. PCG makes heat exchangers and other heavy components at plants in York, Pennsylvania, and Florence, New Jersey, but it can’t make the largest parts, like an AP1000 reactor pressure vessel, CEO Rex Geveden said on the Q2 call. So $BWXT is weighing New Jersey, Indiana, and one other location for a new U.S. large-component plant, and Geveden expects “a final investment decision in the coming months.” He also expected at least one new-build equipment order by the end of 2026, from the three additional small modular reactors (SMRs) planned at Darlington in Ontario, the AP1000 projects, or BWRX-300 SMR projects in the U.S. John MacQuarrie, president of Commercial Operations, widened that window at $BWXT’s Investor Day on September 29, 2026, saying awards should start “either later this year or early in 2027.” $BWXT’s Investor Day targets call for high-teens annual revenue growth in the commercial segment through 2030, with adjusted EBITDA margins rising to the high teens from 13% in 2026. Reading “high teens” as 16-19%, commercial revenue would grow from ~$1.1B in 2026 to ~$2.0-2.2B by 2030, and adjusted EBITDA from ~$140M to ~$320-420M. That would make commercial about a third of $BWXT’s adjusted EBITDA, up from about a fifth.

@convequity Oct 11 4 clicks

Convequity on X

Neoclouds trade at 2.6x invested capital, less than half the hyperscalers' 6.1x The seven neoclouds in Convequity's AI Bubble Barometer (ABBX), $CRWV, $NBIS, $IREN, $APLD, $CORZ, $CIFR and $BTDR, lost 7.2% of their combined enterprise value this week, from $190.2B at the 2 Oct close to $176.5B at the 9 Oct close. Over the same week the five hyperscalers, $MSFT, $AMZN, $GOOGL, $META and $ORCL, gained 2.4%, from $13,316B to $13,637B. 1. Enterprise value divided by invested capital (EV/IC) at the 9 Oct close was 2.6x for the neoclouds and 6.1x for the hyperscalers. At the 2 Oct close it was 2.8x and 6.0x. 2. ABBX estimates the weighted average cost of capital (WACC) at 14.6% for the neoclouds and 9.7% for the hyperscalers. 3. At the 9 Oct close, neocloud share prices were 20% to 45% below each stock's best Friday close since 17 Jul: CIFR −44.5%, CORZ −33.3%, BTDR −25.8%, IREN −24.6%, APLD −24.1%, CRWV −22.0% and NBIS −20.4%. Among the hyperscalers, ORCL was 10.9% below its best, AMZN and META 4.4% and GOOGL 1.3%, while MSFT closed above its best. We think most of the discount reflects the neoclouds' higher cost of capital rather than a mispricing. We would not call 2.6x cheap until they report their July to September backlog and customer prepayment figures.