Bullpen.News · the best place to find stock ideas

@ACapitalLP Oct 6

A Capital on X

@SurfShopCapital $ASTS 🚀

@ACapitalLP Oct 6

A Capital on X

$FSLR <$200 is as dislocated as $TLN was at $280… This is perception <> reality. Do you see it?

@convequity Oct 6

Convequity on X

OpenAI’s TPU lesson makes me more cautious on $CBRS. Richard Ho, OpenAI’s hardware chief and a former TPU designer, says one lesson from Google was that “you have to make it programmable.” Models change too quickly to overfit the chip to one architecture. That caught my attention because TPU is already programmable. Groq statically schedules execution in its compiler, while Cerebras makes a different bet on wafer-scale compute and distributed local memory. The question is how much of the performance advantage survives when models change, including MoE routing and memory demands. Ho also wants hardware that can move across inference workloads without locking the fleet into fixed ratios of specialized machines. That matters when buyers are committing years of capex and power. If OpenAI and other large buyers increasingly favor that flexibility, I expect CBRS to remain a specialist over the next few years. The absolute market can still be large enough for Cerebras to build a very good business, but I’m more skeptical that its architecture becomes the mainstream choice. "As long as you’re willing to live with those limitations, they’re good. But if you’re able to make it programmable, then you have a lot of flexibility for adapting to change. And we had to adapt to change. TPU also adapted to different changes in the algorithm, even at that time with CNNs. Clearly now, in the world of LLMs and agents, the mix is changing pretty rapidly. So that was the pitfall we really wanted to avoid: to overdesign and overfit to one particular model. We looked at it in that context." ---[54] The Man Who Built OpenAI's First Chip TechTechPotato

@convequity Oct 6

Convequity on X

Despite investor doubts, $IREN is worth a small position. The shares have fallen about 17% from their 22 Sep close. The equipment IREN had ordered at 30 June is covered by the cash, committed GPU financing and prepayments it reported on 27 Aug. In the June 2026 quarter customers funded most of its capex. On signed contracts the shares are priced about where $CRWV is. The main risk is that IREN runs GPU clouds worse than its peers. 1. IREN guided FY27 capex (Jul 2026–Jun 2027) at $25–30B on its 27 Aug results call. Its enterprise value is $14.6B at the 5 Oct close. IREN's FY26 results show revenue of $0.7B, of which $0.58B was bitcoin mining. IREN expects to have shut mining down by the end of December 2026. 2. At 30 Jun 2026 IREN had $13.8B of commitments for GPUs and site work, almost all payable by 30 Jun 2027, per its annual report. Orders placed after 30 June are not in that figure. On 27 Aug IREN put its existing cash, committed GPU financing and customer prepayments at about $14B. That covers the 30 June commitments. 3. The other $11–16B of the capex guide is not yet funded. Management targets about $8B more in GPU financing and prepayments for new GPU capacity, and that funding comes with new customer contracts. IREN says it is in late-stage talks for a large share of 2027 capacity. It has not announced any signed contract for that new 2027 capacity. The rest would come from data-center debt, operating cash flow and corporate funding, which includes selling shares. By 14 Aug 2026 IREN had sold about $2.5B of shares under its $6B ATM programme, leaving about $3.5B at that date, per its annual report. It has not reported any sale since. 4. On signed contracts, IREN is priced close to CRWV. On our ABBX dashboard, enterprise value is 0.88x contract value for IREN, 0.74x for CRWV and 1.60x for $NBIS. The contract values are company-reported, and each company measures them differently. IREN reports about $16.6B, of which about $5.1B is cloud service contracts and about $11.4B is GPU lease contracts. CRWV reports a $104.2B revenue backlog. NBIS reports $37.5B of remaining performance obligations. On 27 Aug IREN said its 2026 capacity was largely sold out. It had $4B of contracted annualised revenue run-rate (ARR) for 2026 capacity, which it earns once that capacity is delivered and accepted. $1B of ARR was already operating. EV is about 3.7x the contracted $4B. 5. In the June 2026 quarter, IREN's customers funded most of its capex. On our ABBX dashboard, customer prepayments and deferred revenue covered 87% of IREN's cash capex in that quarter ($1.72B of $1.98B). The same measure was 21% at NBIS and 12% at CRWV. Working-capital inflows covered 90% of IREN's capex and 36% of NBIS's. CRWV's working capital was an outflow equal to 8% of its capex. More than half of IREN's customer prepayments came under the $MSFT contract. At 30 Jun 2026, $1.01B of IREN's $1.84B of deferred revenue sat in the subsidiary that holds the GPUs for the MSFT contract, per its annual report. IREN said on 27 Aug that $3.6B of GPU financing plus customer prepayments fund about 96% of the GPU capex for the MSFT contract. IREN also said its recent contracts carry prepayments of 45–55% of GPU capex. So in that quarter customer cash, not new debt or shares, paid for most of IREN's capex. At NBIS and CRWV it paid for a small share. 6. Operations are the weak point. Some people are questioning the quality of the GPU cloud service IREN delivers to end customers. We share the concern, and it is not new to us. We recorded it in our 1Q26 rebalancing meeting and again in 2Q26 and 3Q26. One of our biggest concerns on IREN is that the CEO treats software as easy, low-value work next to power, land and data centres. That includes the orchestration software that schedules customers' jobs and manages the GPUs. We expected this to slow IREN's growth, not to threaten the business, for three reasons. (A) In our view the bottleneck in AI infrastructure today is powered capacity, not the know-how to run GPU fleets. IREN said on 27 Aug that its recent 3-year contracts price above $20M of revenue per MW of computing capacity. It estimates about a 2-year payback on the GPUs and related equipment. It also said its 2026 capacity is largely sold out. (B) Hyperscalers and other bulk buyers of IREN's capacity can manage and optimise GPUs better than IREN can. What they need from IREN is powered GPU data centres delivered on time. MSFT's $9.7B contract is for GB300 GPUs in four 50MW buildings. MSFT accepted the first, Horizon 1, in August 2026. IREN targets Horizons 2–4 in the December quarter. (C) $NVDA is partnering with IREN and is trying to make running large GPU fleets more standard and simpler. In our view NVDA wants standard fleet-management software so that more companies can run its GPUs well, which makes its chips easier to buy and use. On 7 May 2026 NVDA signed a five-year, $3.4B managed GPU cloud contract with IREN that uses Mirantis orchestration software. The same day it announced a partnership to deploy up to 5GW on NVDA's DSX reference designs. DSX is NVDA's standard set of designs and software for building and running AI data centres. In its 31 May 2026 DSX release, NVDA says the DSX OS software handles tasks such as keeping GPUs healthy, scheduling jobs and running several customers on one cluster. The same release names IREN, CRWV and NBIS among the cloud partners deploying core parts of DSX. The risk in (C) is that IREN's CEO may be too optimistic about how capable NVDA will be at this and how soon. 7. IREN is buying software and hiring managers to run its cloud better. It bought Mirantis, a cloud software company with more than 1,500 enterprise customers, on 4 Aug 2026. IREN said at completion that it paid about 12.6M shares plus about $40M. In July 2026 it hired a product chief from $ORCL (its cloud unit), a data-center development chief from $GOOGL and CyrusOne, and a security chief from $NTNX. We found no hire of a head of cloud operations. We would change our view in three cases. First, if a Horizon delivery for MSFT is missed or rejected. Second, if the Q1 FY27 report arrives with no new contract for 2027 capacity announced. Third, if IREN's weak end-user experience turns out to come from its infrastructure, meaning power, cooling, networking, storage or the GPU hardware itself, and not just from its orchestration software. The third matters because bulk buyers can replace IREN's software with their own, but they depend on IREN for the site and the hardware.

@convequity Oct 6

Convequity on X

Reflection AI has committed to more than $7bn of compute, and it will pay for it with investors' money, not with revenue from its new model. 1. On 5 October Reflection launched Beam, an open-weight model, meaning anyone will be able to download and run it for free. Most of the talk on X is about its benchmark claims, which have not been independently verified. 2. Reflection does not plan to sell access to the model. It plans to make money by helping enterprises and governments train its models on their own data and run them in-house. 3. This summer it agreed to pay $SPCX $150m a month for chips through 2029, about $6.3bn over the full term. Either side can end that deal with 90 days' notice after the first three months. It also signed a deal with $NBIS worth more than $1bn. 4. Reflection has raised about $4.7bn. That cash, not model sales, is what pays these bills. 5. Nebius says prepayments covered 50-60% of the capex on its four big Q2 deals, including Reflection's. It has not disclosed how much of Reflection's own contract is prepaid, guaranteed or cancellable. Until it does, I would haircut Reflection's share of Nebius's backlog. Reflection is a small part of Nebius's $40bn+ of customer commitments, so the effect on the total is small. ABBX, our AI bubble barometer, says the AI market is not in bubble territory yet: forward returns on AI capital are still above the cost of capital. It names commitments that do not show up as debt as the main risk. A lab with no model revenue signing multi-year compute contracts paid for with venture money is that kind of commitment. If more deals like this appear, I will read them as an early sign the market is getting frothy.

@DeepSailCapital Oct 5 2 clicks

Deep Sail Capital on X

$MELI - Long term vision. https://t.co/EXtoxpWBg3

@RagingVentures Oct 5 1 click

Raging Capital Ventures on X

@TarekAntaki I bought $CHTR stock today as well, Godspeed!

@ACapitalLP Oct 5

A Capital on X

𝗣𝗲𝗿𝗰𝗲𝗽𝘁𝗶𝗼𝗻 𝗜𝘀 𝘁𝗵𝗲 𝗧𝗿𝗮𝗱𝗲. It’s funny: most people think the proliferation of information has made finding dislocations harder. I’d argue the opposite. When information was scarce, speed was the edge. Today the problem is information overload. Social media, AI, headlines, experts, hot takes. Just like different news organizations (CNN, Fox, MSNOW) can frame the same political candidate through completely different lenses, markets create narratives that can overwhelm the underlying facts. Every trade is a spread between perception and truth. Merger arb just makes that spread explicit because you have a relatively fixed endpoint (the deal consideration). Regular-way equities are harder because “truth” is a moving target. But you don’t need to know whether $TLN is worth $350, $370 or $390 to recognize when Josh Shapiro and data-center pushback have distorted perception enough to send it to $270. You don’t need to perfectly value every business line of $OWL to recognize when a Bloomberg story has caused the market to paint the entire business with the private-credit apocalypse brush. And it works both ways: $GLND wasn’t suddenly a critical-minerals bonanza because of a Trump headline. Sometimes the narrative creates the overvaluation. Having the best model or talking to the most experts helps you estimate truth. But truth without an understanding of perception isn’t a trade. You can be a little wrong on intrinsic value and still make a lot of money if you correctly identify that the market’s narrative has overshot reality. The edge today isn’t knowing whether something is worth exactly $37 or $39. It’s understanding why it’s trading at $27 in the first place, whether that perception is wrong, and what makes the gap close…