Convequity

Everything we have picked from this writer, newest first.

@convequity Jul 29

Convequity on X

Ciena — the underappreciated coherent layer As AI clusters move beyond a single building into campus and multi-site “scale-across” architectures, short-reach optics hit their physical limits. Signal distortion and dispersion make conventional detection unusable over distance. $CIEN’s coherent technology recovers both amplitude and phase, reconstructing a clean signal where others cannot. The market remains focused on the high-volume short-reach stack — pluggables, CPO, and related optics inside the data center. Coherent is the quieter, longer-reach piece that becomes essential once clusters themselves become distributed.

@convequity Jul 28

Convequity on X

Kimi K3 open source just made local frontier inference a different game 1.56TB download. 2.88T total / 104B active. Recommended: 64-chip supernode. Individual users are effectively locked out of running the new frontier open-source models at home. The only realistic path left is hoping Apple’s M7 Ultra with Thunderbolt-based RDMA can handle it — and that still depends on memory pricing from Samsung/SK Hynix/Micron or whether Washington lets Apple source from CXMT/YMTC. The bigger shift is on the enterprise side. If you run K3 on an NVL72 rack without stuffing the batch with enough concurrent users, your inference cost can easily run 5x+ versus an API provider that aggregates queries. Most companies will need an orchestration layer that either: - pools multiple user requests in real time, or - parks non-urgent demand for overnight runs to maximize hardware utilization. That reality makes GPU rental structurally more attractive than buying and operating the iron yourself. Which is an uncomfortable implication for $DELL’s enterprise GPU server business. Curious how many enterprises will actually choose to run models at this scale on-prem without sophisticated batching. The economics are no longer obvious.

@convequity Jul 28

Convequity on X

Rocket Lab just won its biggest launch contract ever: $266 million from the U.S. Space Force The deal covers 12 guaranteed suborbital test launches (with options for 6 more) for missile-defense work. First flight no earlier than the end of 2026, mostly from a brand-new Rocket Lab site in Alaska. This locks in real multi-year defense work and a new U.S. launch site. Solid backlog signal for $RKLB — and exactly why we keep the position in the Convequity portfolio.

@convequity Jul 28

Convequity on X

Samsara and the physical-AI narrative We exited $IOT. The market is still pricing it as a core beneficiary of the coming wave of AI-enabled robots in warehouses, offices, and homes. Our read is different. Samsara’s real differentiation is software orchestration for distributed IoT fleets, not proprietary sensors or hardware. The physical sensors are outsourced to Asian manufacturers, and the orchestration layer itself is exactly the part of the stack that AI coding tools are making easier to replicate. As a result, the competitive position looks less durable than the current narrative suggests.

@convequity Jul 28

Convequity on X

Rubrik’s quiet edge in the agentic era As enterprises deploy autonomous AI agents, the real constraint is becoming governed access to trusted data — not just more models. $RBRK’s Annapurna is building exactly that layer: indexing unstructured data in place, preserving Zero Trust controls, and turning the backup estate into a secure, queryable foundation for AI. In a world where agents can move fast and break things, the company that already owns the governed copy of enterprise data is well positioned.

@convequity Jul 28

Convequity on X

The complacency is the point. Gas-directed rigs are still only in the low-to-mid 120s and the trend remains essentially flat. Turbine lead times run 3–5 years and the major OEMs are already sold out into 2029–31. Data-center announcements are loud today, but the actual incremental gas draw from behind-the-meter and new generation does not hit the system until the turbines arrive and the sites come online. That lag is why the market can keep treating gas as abundant through 2026–27 even as the structural deficit builds. When the convexity shows up, it will likely arrive later and harder than most are pricing. $EQT, $AR, $CRK

@convequity Jul 24

Convequity on X

NVIDIA’s Feynman: CPO Inside, OCS Outside According to supply-chain sources, $NVDA’s 2028 Feynman architecture is planning a dual-layer optical design around the NVL576 rack. An inner layer of Co-Packaged Optics will handle high-density, short-reach connections within and between racks. An outer layer of Optical Circuit Switching will manage larger-scale cluster connectivity, supporting flexible topologies at the 100,000-GPU level. In this framing, CPO and OCS are treated as complements rather than substitutes. CPO addresses density and power at short reach; OCS provides flexibility at longer distances. NVIDIA’s near-term path appears to favour mature MEMS switches to secure delivery schedules, with silicon photonics as the longer-term option once insertion loss improves. Higher insertion loss in SiPh currently requires more optical amplification, adding power, noise and cost — which is why MEMS remains the lower-risk choice for large deployments in the near term. $GOOGL’s earlier move into silicon photonics for inference is expected to help mature the supply chain, potentially making it easier for NVIDIA to follow later. Huawei, Oracle and others are also advancing along related paths.

@convequity Jul 24

Convequity on X

Photonics Pt.4: OCS Roadmaps Are Splitting Convequity’s latest note examines how Optical Circuit Switching is diverging by workload. High-radix MEMS remains the practical choice for scale-out and training clusters that prioritise scheduled reconfiguration. Lower-latency silicon photonics is being pushed into the scale-up layer for inference, where millisecond switching is no longer acceptable. $GOOGL's V8i architecture is the clearest example. It moves OCS down into the scale-up domain for the first time and pairs it with 2.4T Coherent Lite modules that dramatically reduce fiber and port consumption. The combination is designed to close Google’s main networking gap versus NVIDIA. $NVDA's Feynman (2028) appears to be taking a dual approach: CPO for dense short-reach connections and OCS for larger cluster-level scale-out. Near-term deployments will likely rely on mature MEMS, with silicon photonics as the longer-term path once insertion loss improves. The broader OCS market is projected to grow from ~$400M in 2025 to over $2.5B by 2029 (58% CAGR). Silicon photonics offers a relatively open competitive field compared with MEMS, where incumbents hold stronger positions. We also re-examine CPO’s path forward and assess microLED as a more radical parallel alternative — one that inverts the industry’s “fewer, faster lanes” assumption and is physically confined to short-reach scale-up links. Full Part 4 is now live. Link in the comments.

@convequity Jul 22

Convequity on X

32% short interest on $SPCX is crazy. Although some AI bulls are insane & unsophisticated as well, short sellers of SpaceX over the longer term is even more out of the rail. https://t.co/Ys2mb3B6oC

@convequity Jul 20

Convequity on X

Crypto Basket Update: COIN, CRCL & HOOD Within our “buy the dip” crypto basket we are de-prioritizing Coinbase relative to Circle. $COIN lagged amid prolonged CLARITY Act uncertainty and faces revenue risk from the ongoing USDC custody renegotiation with $CRCL. Broader Bitcoin structural concerns (miners shifting to AI) also temper the trading-volume outlook. Circle stands out as the preferred stablecoin exposure. Stablecoins are the natural token/payment rail for AI agents — a structural driver distinct from trading hype — with added tailwinds from interest rate sensitivity and attractive asymmetry after the pullback. $HOOD is held at lower conviction as one of several interchangeable infrastructure bets. We are maintaining current weights across the basket but leaning toward adding to Circle as the highest-conviction name.

@convequity Jul 20

Convequity on X

Palantir remains our marquee software conviction. Growth is accelerating as ontology and governance become essential for secure AI deployment in government and enterprise. The bear case is model dependency, but this is where $PLTR shines. Enterprises risk losing their information and knowhow edge by feeding proprietary data to foundation model providers — their own suppliers. Satya Nadella’s “Reverse Information Paradox” captures it perfectly. Microsoft did it with Lotus/WordPerfect (Excel/Word), and Anthropic is already demonstrating such practices with Claude Design competing with Figma — and countless other examples too. Palantir’s strong data control and governance moat positions it to avoid the trap. Developing its own foundation model could be a game-changer, offering true IP protection.