Saturday, November 7, 2026

@leevalueroach Oct 19 1 click

Lee Roach on X

Geo Group $GEO sold its 3 Adelanto ICE facilities (2,644 beds) to DHS for $950M, about $359K per bed. The market currently values GEO at about $92K of EV per remaining owned bed, and that EV also includes the electronic monitoring and services businesses.

@finphysnerd Oct 19 1 click

Myles on X

If you're naturally bearish like me I think it pays to push long when positioning gets washed out like it is now even if fundamentals are bad. Retail has been selling, hedge fund net leverage is low, $RSP and $QQQ have been smashed. You are going to be right 60-80% of the time,

@finphysnerd Oct 19

Myles on X

Closing my $QQQ and $SMH shorts here and lowering my $IWM exposure. I think we're bottoming here so keeping net long exposure, but the Q's and tech have really lagged over the last 2 weeks. Very happy with this trade.

@StableBread Oct 19 11 clicks

Fajasy on X

I first wrote up an AI data center company that rents out $NVDA GPUs to AI startup companies on April 15, 2026, when it still traded as a SPAC at $10.70/share, just above its ~$10.60 trust value. The merger was a success, and the stock peaked at $39.20/share on June 1st. Now it trades below its SPAC price. That company is Boost Run $BRUN, which has fallen to $8.89/share, down 77.3% from its peak, even after it announced its largest disclosed contract on October 6 (worth $525.6M). The three larger listed GPU clouds have fallen from their 2026 highs as well, through the October 16 close: → CoreWeave $CRWV: Down 55.5% to $61.39/share. → $IREN: Down 60.2% to $26.99/share. → Nebius $NBIS: Down 42.3% to $165.47/share. So $BRUN fell with the group, but further than all three. I’d blame the drop from $39.20 on two causes that have nothing to do with $BRUN’s core operations: (1) higher rates and borrowing costs, and (2) $BRUN’s float growing ~3.1x, which was expected once its lockup ended and its warrants were called. Yet the business carries less risk than it did as a SPAC: → Contracts: $1.61B with disclosed terms and TCV over $2.6B, including $1.43B signed since my April write-up. → Customers: Thinking Machines Lab (the AI startup founded by former OpenAI CTO Mira Murati) and Cohere make up 61.9% of the $1.61B on three- to five-year terms, after three customers made up 75.9% of 2025 revenue, led by RunPod, a GPU cloud platform, at 44.5%. → Year-end target: If Thinking Machines’ GPUs are fully running by December 31 and the June order starts on November 30, the five disclosed contracts running by year-end cover $300.6M (75.1%) of the $400M ARR target. → Merger: Closed on May 8, 2026, with zero redemptions. → Balance sheet: The merger proceeds paid off its short-term bridge loans, the going-concern language is gone, and pro forma cash is $191.8M, up from the $108M I used in my June 3 update, right after the merger. But $BRUN still has to get the GPUs bought, financed, and running by December 31: → To reach the $1.0-1.2B of GPU spending management expects for 2026, $BRUN has to deploy $663-863M of GPU servers in H2 2026, 2.3-3.0x its cash and contracted prepayments. → CFO Erik Guckel said on the Q2 call that $BRUN is working on new ways to finance the build, but I found no new facility, lender, or term sheet in its filings or press releases from August 14 to October 16. → $BRUN’s named customers are private AI startups without investment-grade ratings, so its contracts are harder to borrow against. → If lessors funded 70% of the H2 2026 build instead of 79.5% (as in H1 2026), $BRUN would need $199-259M of its own cash for the build, more than its $191.8M of pro forma cash. → Both 10-Qs came late, after notices of late filing on May 26 and August 14, and $BRUN still reports three material weaknesses, the most serious kind of gap in a company’s financial reporting controls. → Founder and CEO Andrew Karos controls 87% of the vote. At $8.89/share, $BRUN’s EV is $989.7M. That’s 2.5x its $400M target for year-end ARR, down from 5.4x at the $22.93 price in my August 16 update. At 2.5x, $BRUN trades below all three peers, including $IREN at 3.2x. So did the thesis break, or just the stock? That’s what I answer in my new ~5k-word update on $BRUN, including my bear, base, and bull cases (Excel model included).

@ToddWenning Oct 19 13 clicks

Todd Wenning on X

One of the big knocks on the thesis is that $HD or $LOW would copy whatever they're doing. HD had the chance back in the mid-2000s when Howden Millwork was forced to retreat, but didn't. $FND also built a multi-billion enterprise under their noses. The key is to focus on a niche and then be really good at that niche, which I think HWDN can do.

@convequity Oct 19

Convequity on X

We think the foundry business at Intel ($INTC) may be closer than the market assumes. In our view, the key question is whether outside customers can design and produce chips with Intel. That is hard because Intel built its design rules, software and processes for its own teams. Two developments help. 1) Intel's 12nm partnership with $UMC brings experience serving external customers and a more customer-friendly design kit. 2) Musk has said Terafab plans to use Intel's 14A process, with Tesla and SpaceX building and running the fabs themselves. We see that as an important external test of 14A. Intel still has to prove yields, volume and schedules, and Terafab's commercial terms remain unclear. We continue to hold our Intel position and remain bullish, but not enough to add to it for now.

@KEDM_COM Oct 19 22 clicks

KEDM.com on X

“There's only 10 ways to do fraud.” @hkuppy and @herbgreenberg on why the same patterns keep showing up, and the thin line between a visionary and a fraud. Carvana $CVNA is a good example. A few years ago, it looked like the company might run out of money, with Herb joking that its car vending machines could become trampoline parks or indoor skydiving towers. Instead, the stock recovered and access to capital reopened. Being right about a questionable business doesn’t help much if it can keep financing itself.

@convequity Oct 19 7 clicks

Convequity on X

Nvidia ($NVDA) faces a major issue: most AI compute is moving to inference, where its CUDA moat protects less. CUDA's programming flexibility is essential for training, which is complex and changes rapidly. Inference runs a trained model with fixed weights, and we think inference chips are easier to build than training chips. The mix is shifting. Post-training RL, reasoning tokens and final replies are all forms of inference. On our view, pre-training could fall below 20% of compute in the future. In February, Cursor said the compute used to post-train its Composer 1.5 model surpassed the compute used to pre-train the base model. At Convequity we think CUDA's depth now protects a shrinking share of the workload, and we expect that to continue. Deeper dive available at Convequity.