Bullpen.News From Edwin Dorsey
Get Daily Recap Feedback

Monday, November 9, 2026

Back to the front page

Roblox: Capture This Moment

Roblox’s Q3 FY26 DAUs fell year over year for the first time, while bookings declined 12% to $1.7 billion.

RBLX $35B
TSOH Investment Research 1 click

From X

@convequity 1 click

Convequity on X

$SPCX negotiated a 90-day right to walk away from Anthropic. We infer that Musk expects Grok to catch up with Claude, and that SpaceX would then cancel and use the Anthropic compute for Grok rather than keep serving Claude. Either party holds that right. Anthropic's S-1 says each agreement can be canceled by either party on 90 days' written notice after a period when cancellation is not allowed. Those periods end in January 2027 and March 2027, depending on the agreement. Anthropic's S-1 says Anthropic agreed to spend up to $120.5 billion into 2029 with SPCX, for compute powered by $NVDA GPUs. SpaceX's S-1 says either party can terminate on 90 days' notice after an initial three-month period. SpaceX told the SEC it negotiated the clause so the capacity could go to other customers or to SpaceX's own compute initiatives. The Apollo finance lease does not involve Anthropic. SpaceX has a binding commitment to enter a separate lease of about $40 billion for Nvidia Rubin GPUs. The 8-K and the press release do not name Anthropic.
@convequity 2 clicks

Convequity on X

Investors are selling $BESI, in our view, because customers increasingly prefer or consider shorter HBM stacks, since HBM is expensive. Our industry information points to customers requesting stacks of four or six memory dies. Fewer layers mean fewer connections and, where hybrid bonding is used, fewer bonding steps per stack, and shorter stacks fit package-height limits more easily with conventional solder. That could delay equipment adoption and weaken the nearer-term HBM opportunity for $BESI, which supplies hybrid-bonding equipment. Logic folding is the tailwind. It places parts of a circuit above one another on vertically connected layers joined through hybrid bonding, so signals travel shorter distances and the chip runs faster on less energy. Those benefits hold even when an inference system needs less HBM, and the layers still need aligning and connecting with exceptional precision. In our view, that can create more hybrid-bonding demand than this delay takes away. Wider folding also pushes inspection and measurement into more logic-chip applications. We expect that to lift spending faster than chip volumes for $ONTO and $CAMT, and for $NVMI, whose materials and chemistry checks address problems that get harder as circuits move into vertically connected layers. Longer term, folding could make chips on established processes more capable for $UMC. That is our prospective thesis, not an announced UMC program. Deeper dive available at Convequity.
@convequity 1 click

Convequity on X

$AVGO could find it worthwhile to stand behind the chips in Anthropic's $63bn financing if the chip sales it enables are profitable enough. On October 26, $BX and three banks announced $63bn of committed financing for a new special purpose vehicle. The vehicle buys AVGO AI accelerators and networking equipment and leases them to Anthropic. Deliveries are expected from calendar 2027 under a compute agreement AVGO had already announced. The debt has two layers. First-lien lenders have the first claim on the vehicle's assets if it defaults. Second-lien lenders are paid from whatever remains after them. The banks arrange the first lien. BX is the sole provider of the committed portion of the second lien. AVGO provides residual value support to the vehicle for an amount equivalent to the first-lien portion, meaning it stands behind the future value of the equipment for that amount. The release gives no loan split and no payment terms. Reuters, citing two people familiar with the deal, puts the first lien at about $43.5bn, or 69% of the total. The support is a promise about the equipment's future value, and the conditions under which AVGO would pay are not disclosed. Its economic cost to AVGO is a weighing of how likely a payment is and how large it could be. A second factor is the cash or borrowing capacity AVGO has to keep available to honor the promise while it stands. If the promise lets Anthropic finance chips it would otherwise have delayed or not bought, and if those chips sell at a profit, the gross profit on them belongs to AVGO. Gross profit means the sale price less the cost of making the chip, before research and other operating costs. Whether any sales are extra in that sense is not disclosed, and neither is the deal's gross profit. That comparison explains why a chip vendor could take a risk a lender might refuse. A lender earns interest and fees on the loan. AVGO could earn the profit on the extra chips as well, and if that profit outweighs the likely payments and the cost of keeping capacity available, the support adds value for AVGO. The weak point is that the costs and the profit move together. The equipment loses resale value when demand for AI accelerators falls, and AVGO's own chip sales would fall in the same conditions. BX's second lien sits below the senior layer that AVGO's support is sized to. BX is therefore lending mainly against Anthropic's ability to pay its leases. Reuters reported on October 16, citing the Financial Times, that Anthropic was then offering BX a yield above 10% on $9bn of debt backed by five-year lease payments. Reuters could not verify that report, and the final yield is undisclosed. If other lenders are charging more for AI lease risk, or declining it, BX gets paid for taking it. Whether that pay covers the risk depends on lease terms that are not public. The same reasoning can apply further up the supply chain. A supplier could stand behind the value of its own equipment if its gross margin on a deployment is large enough to cover the likely payments, it has the capital to keep available, and the support terms are ones a lender can enforce. If credit for AI data centers stays tight, that is one way the buildout keeps getting financed, with suppliers carrying part of the risk that their own hardware loses value.
@convequity 1 click

Convequity on X

At similar valuations we prefer $AVGO over $NVDA, because Broadcom has more room to beat expectations. A custom-chip order still needs lasers, optical and electronic chips, and packaging, and Broadcom supplies all of them. Separate suppliers each protect their own margin, so cutting total system cost is harder. Broadcom could also win custom-chip orders by assuring customers that the optical components needed to deploy them will be available. Google TPU volume adds a second large revenue source. Estimates put Google's TPU-related contribution at roughly $11 billion to $13 billion in 2025. Supply-chain estimates put Google's TPU v8 shipments at about 4.0 million units in 2026, rising to about 9.1 million in 2027. Assuming Broadcom supplies the remainder after MediaTek's allocation, that leaves a large expansion for Broadcom. Nvidia starts with a much larger share, so more of its business is exposed as customers develop alternatives.
@ACapitalLP 3 clicks

A Capital on X

Interesting special situation in $AKER.NO, which owns ~23% of Nscale. Despite the carnage in neoclouds, the FT reported last week that Nscale may still be targeting a $30–35bn IPO before year-end. What makes this interesting is the valuation disconnect. At ~NOK 1,150, backing out Aker’s other holdings and applying its historical ~22% HoldCo discount implies an ~$11bn valuation for Nscale. Below its $14.6bn March round and well below the proposed IPO range. In September, Third Point led a $3.36bn pre-IPO convertible financing alongside Nvidia, Apollo, Citadel and others. The notes reportedly convert at a double-digit discount to the IPO price, with a $30bn valuation cap on the conversion price. Not exactly a group of unsophisticated investors… At a $25–30bn IPO, I get ~27–39% upside assuming some HoldCo discount compression. Even at a $15bn IPO, the downside looks relatively contained under a historical discount assumption. Obviously the IPO could slip, financing and execution risks are enormous, and the broader neocloud unwind isn’t helping. But you can partially hedge the sector exposure with $CRWV / $NBIS shorts, leaving a more isolated bet on the IPO valuation and HoldCo discount...

Investor Letters

All letters

NEW POSITION: Google LEAPS

Alphabet bought long-dated GOOGL calls after a 12.6% earnings drop, betting AI monetization and distribution overcome capex pressure and…

GOOGL $3.5T
Hated Moats 1 click

Axon 3Q26 Business Update

Axon cut full-year guidance after delayed federal Dedrone contracts slowed 3Q revenue growth to 28%, while software and services grew 43%.

AXON $31B
Speedwell Memos 4 clicks

AMD Q3 FY26, Annotated

AMD's Q3 revenue rose 45% to $13.4 billion as data center reached $8.5 billion, 64% of sales, driven by EPYC and Instinct ramps.

AMD $660B
Chipstrat 2 clicks

Publications we follow

Every publication on this site, hand-picked.

129 publications

Newsletters 109

YouTube 4

Blogs 15

Seeking Alpha 77 of 280

Nominate a publication

X accounts we follow

Every account on this site, hand-picked.

91 accounts
Nominate an account

Investor letters we follow

359 firms

Sectors

Ordered by how often our editor reaches for them.