Saturday, October 24, 2026

Back to today

Tower Semiconductor Deep Dive

Tower Semiconductor’s low-loss photonics chips halve lasers in 1.6T optical modules, supporting a $680 million annualized photonics run rate and customer wafer prepayments through 2027.

From X

@convequity

Convequity on X

$RBRK's backup engine breaks data into fixed-size blocks, for example 4 MB, and stores only the unique blocks across backups. Commvault and Cohesity instead divide data into variable-sized chunks, with the boundaries set by the content itself. Each design has a trade-off. With fixed blocks, a small change inside a block means the whole block gets written again, so Rubrik uses more storage. In return, hashing and comparing blocks is simple, which means faster backups and lower compute overhead. With variable chunks, only the changed portions are written, so storage use is much lower. Working out chunk boundaries and hashes as the data comes in takes more CPU and memory, which can slowdown backups. Cohesity offsets part of this with a scale-out architecture that processes in parallel across nodes. The two designs fit two different jobs. Variable-length deduplication suits long-term archives, where storage cost matters more than speed. Fixed-block deduplication suits recovery after a hack or ransomware attack, where fast restores and predictable performance matter more than saving storage. That matches how each company started. Rubrik came into the market in 2014 focused on security and on getting critical data back quickly after a breach. Cohesity started from a different idea. It built its own file system that spreads data across many servers, each with its own computing power. Backup copies stay readable like normal files, so companies could search and analyze them where they sit, without restoring them first. Security came second. The two are now moving closer together. Cohesity has leaned much more into security and recovery, especially after buying Veritas's backup business in 2024.
@convequity

Convequity on X

$NTSK began as a CASB vendor, governing how employees use SaaS applications at the point of use. Years of that work gave it application-specific knowledge of user behavior, data movement, and risk inside cloud apps. That knowledge carries over to AI. Netskope inspects prompts, responses, and file transfers inline and in real time, and it can respond with something more precise than a block: preventing a risky copy-and-paste, stepping up authentication for a sensitive interaction, or coaching the user as they work. The part that matters more comes after the prompt. AI adds autonomy and chaining, so a single human interaction can trigger many machine-executed actions across tools, data stores, and services, often outside one user session. If data is over-permissioned and tools are loosely scoped, an agent can stay inside its allowed permissions and still cause harm faster than a person can intervene. The context Netskope gathers at the interaction can be carried forward to constrain that downstream execution: who the user is, what data is involved, which action was taken, and how risky it looks. So Netskope's edge is that it already knows the user, the data and the action at the moment of the prompt, and it can use that context to limit what an agent does next. It is now building the agent side directly. Its Agentic Broker rates agent actions by risk before they run, and its AI posture management for agents built on AWS, Azure and GCP is still in preview. That is the part to watch.
@convequity 2 clicks

Convequity on X

More $NVDA chips could end up going to SpaceX and other buyers that can borrow at investment-grade rates, leaving neoclouds with fewer chips than they need. SpaceX's planned $40bn chip borrowing shows why. The deal: $SPCX is in early talks for about $10bn of bank loans and $30bn of investment-grade debt, led by Apollo, with Pimco also in talks. It would close in 2027, and the talks could still fall apart. The comparison: in December, Apollo lent about $3.5bn to fund xAI's Colossus 2 chips through a separate company set up just for that deal. The interest rate was about 10% a year. xAI is now part of SpaceX. Investment-grade debt usually costs much less. US investment-grade corporate bonds yield around 6% on average today. The rate on the new $30bn has not been reported, so we don't know the exact saving yet. Why it could be cheaper: lenders would be relying on SpaceX's launch and Starlink businesses to repay, not only on the chips. That is the real financial payoff of putting xAI inside SpaceX. Most GPU buyers can't borrow like that. $CRWV is rated junk (Moody's Ba3, Fitch BB-) and pays 9.25% on its 2030 bonds. Standalone AI labs have no public credit ratings that we could find. Not every neocloud pays up. $NBIS has a GPU loan at a floating benchmark rate plus 2.5%, against CoreWeave's plus 5.5%. Our view: the debate on X is whether this proves AI demand or shows too much leverage. We think the bigger point is cost of capital. When chips are scarce, the buyer that borrows most cheaply can pay for the most of them, and neoclouds paying junk-level rates get squeezed out.

Our Views on CoreWeave

CoreWeave's financial statements point to a business model familiar to the authors after a closer review of the AI infrastructure company.

Capital Wars

CoreWeave’s $35 billion to $39 billion 2026 spending plan relies on debt as unsecured yields reach 15%, making its GPU cloud expansion a…

Putting bear cases to rest

Medpace authorized $1 billion of buybacks through 2028 after Q3 revenue rose 8.5% to $716 million; service revenue grew roughly 12% as…

EXPE · long

Expedia Group INC

Expedia Group, one of the largest online travel companies, has roughly 3.6 million lodging properties and is expanding internationally.

Publications we follow

Every publication on this site, hand-picked.

128 publications

Newsletters 108

X accounts we follow

Every account on this site, hand-picked.

91 accounts

Investor letters we follow

190 firms

Sectors

Ordered by how often our editor reaches for them.