Anduril's Thunder launch, arriving one week after its AWS partnership to push data center capability to the battlefield, matters less as a single product than as evidence of release cadence. The valuation framework here is straightforward. Raytheon carries a $261 billion market cap as the closest thing to a pure play defense prime, with Boeing at $166 billion, Lockheed Martin at $118 billion, General Dynamics at $100 billion, Northrop Grumman at $74 billion, and L3Harris at $52 billion. Anduril's implied valuation sits near $100 billion, which frames a $225 billion to $275 billion resting range post IPO if the company continues to serve sovereign militaries exclusively. That is roughly 2.25x to 2.75x from current marks over a two year horizon, and few investors would object to that outcome. The larger optionality is commercial. Raytheon finished 2025 on $88 billion of revenue, and public markets do not award hardware businesses generous multiples, whereas Palantir reached roughly 60x revenue precisely because it opened a commercial TAM (total addressable market). Management has not signaled a commercial pivot, though counter drone and autonomous surveillance for stadiums and corporate real estate are the obvious adjacencies, and demand is building from Ukraine, Israel, and Southeast Asia for frontier systems on frontier timelines. We remain constructive on Anduril and considerably more cautious on single product defense names, where selling one weapon system to a finite set of militaries caps growth and tends to end in acquisition by a prime rather than durable public compounding.
Kimi K3 clearing frontier benchmarks on open weights is the most consequential AI development of the month, and it is being widely misread. Running a Chinese open weight model does not route data to China, because open weight models are files that are downloaded, run on local infrastructure, and send nothing back. The more useful question is why the best open weight models are Chinese, and the answer is constraint rather than talent alone. Chinese labs have been building without unrestricted GPU access, which forced software side innovation such as agent orchestration and dynamic routing to smaller specialized models, techniques that were novel six to seven months ago and are now standard practice. We expect 80% to 85% of global AI compute spend to run on open weights within five to ten years, against a current closed frontier share near 75% to 80%. That is not a bearish call on Anthropic. Compression to a 15% to 20% share of a pie growing exponentially still implies roughly 10x current revenue, and the reported October IPO timeline is the nearer catalyst: the business was at $4 billion of revenue in December against a run rate now reported near $100 billion, which likely prints closer to $40 billion to $50 billion on a GAAP quarterly basis. No frontier model provider has come public yet, with CoreWeave being infrastructure rather than a model business, so investors will see that revenue slope alongside a Fortune 500 client roster with no forward guidance permitted in the filing. We expect a strong debut with the usual post IPO volatility, and we would treat open weight competition as a narrative risk rather than a revenue risk.
The Figma episode is the clearest available lesson in platform risk. Figma embedded its design tooling into Claude, and Anthropic subsequently launched Claude Design into the same category, which is ordinary platform behavior and precisely what shareholders should expect from an 800 pound gorilla. The investable conclusion is to own the layer beneath the AI platforms - AI infrastructure companies - and be very selective in the application layer sitting on top of AI platforms. Whether inference runs on a closed frontier model or an open weight model fine tuned in house, it runs on semiconductors, data centers, model routers, and inference providers, and names such as Baseten, Fireworks AI, and Factory AI occupy that metal layer at $5 billion to $15 billion valuations, which makes them midcap businesses in everything but listing status. Enterprises increasingly want to run their own model on their own data so that workflows and internal process remain proprietary rather than training a frontier lab's next release, and that behavior expands inference demand regardless of which lab wins. On pre-IPO portfolio construction, we continue to advocate tranching across vintages rather than a single entry, deploying annually and recycling year three to five liquidity into the next tranche, because time diversification matters as much as name diversification when technology cycles move this quickly. Nearly all of this opportunity set is private and largely uncovered by public market analysts, which is the source of the inefficiency.
OpenAI's first hardware product, reported as portable, screenless, and voice native, should not be evaluated as an iPhone competitor. It is a step toward ambient capture, and we expect audio recording of everything a person says or hears to become near universal within 6 to 12 months with video following shortly after. The value is not in the device. It is in the proactive AI layer that sits on a complete personal data record so the AI personal assistant can surface what was missed, what is due, and what to do next. That implies multiple form factors rather than one dominant device, including stationary units in homes and offices, which is a modest extension of the Echo and Sonos install base already in place. The privacy objection is real but has historically not held, because consumers accept data capture when the value exchange is favorable, and enterprises already claim ownership of all employee work product by contract. We expect businesses to be the first adopters, recording meetings, calls, and screens, and eventually training on that corpus to retain institutional knowledge after employees depart. The terminal state is a direct neural interface, already functional in a limited clinical setting via Neuralink, though that is a ten to thirty year arc. For now the investable question is who owns the personal data layer, and Apple, Google, and Amazon remain better positioned than OpenAI and Anthropic to capture it.