CoreWeave's bonds have sold off, with yields approaching 15%, even as the company carries a $104 billion backlog, and we see the disconnect as a dislocation in the market and a buying opportunity. CoreWeave is a neocloud that buys GPUs, stands up data centers and leases the compute, and few banks are willing to treat a GPU as high-quality collateral, so the company is paying up for its debt. A lender that can properly price GPU clusters together with the energy, cooling and full operation behind them, such as Palmer Luckey's new bank Erebor, could collateralize those assets more efficiently and improve the company's investment profile. When a company faces a tidal wave of demand it cannot fully meet and executes on par with or slightly better than its peers, we view that as a strong setup for investment returns. We believe the negativity reflects investors who do not yet believe AI will play out, since most people still use it as a replacement for Google, while AI companies and hyperscalers are seeing a wave of enterprise demand from customers asking for twice the compute they have and unable to find the electricity or GPUs to get it. SpaceX shows the other side of the trade: it is adding billions of dollars of incremental revenue from data center and cloud leasing, can stand up compute faster than anyone, and management pointed to a return on investment of about six months and a revenue run rate of $500 billion to $1 trillion by the end of 2027. Even if SpaceX executes twice as well as CoreWeave, demand is large enough that CoreWeave still looks good … in our opinion.
Modal has reportedly crossed $1 billion in annualized run rate revenue and raised a $750 million Series D at a $15.75 billion post-money valuation led by Accel, and we like both the company and the category it represents. Modal is an AI inference and training platform, part of an emerging group of developer tools that are not data centers or chips but are clearly infrastructure, helping developers customize their own models, run them in a harness and work with open-source models. Its serverless sandboxes let developers run models and code without direct access to a server, which keeps workloads contained, and from what we see Modal is the industry leader in that area. We see a shift away from Anthropic and OpenAI, driven by cost, toward open-source models that businesses and AI application companies customize for better output, and that work requires platforms like Modal rather than in-house builds. Our preference is an allocation to several businesses in the space rather than a single bet, including Modal, Baseten, Together AI and Fireworks AI, with the caveat that some are more expensive than others. The open question is how large these companies can become and whether they can reach $100 billion valuations. Databricks, a developer platform at scale, is valued at more than $200 billion in the secondary market, which makes Modal look attractive at a $15.75 billion post-money valuation.
ElevenLabs has reached a $22 billion valuation and released a real-time voice model, and we see it as an interesting AI application company that has clearly found its market in voice. A year ago the use case was unclear, but the company now serves customer service and other voice applications, and it has added models for music and sound. The new real-time model is designed to remove the lag in current voice assistants and respond with human-like inflection and back-and-forth timing, which matters as more users talk to AI assistants instead of typing. Users can also upload a video, such as a boxing match or a sports game, and the model analyzes the scene and creates matching sound. Advertising companies and creators are reportedly using ElevenLabs to produce their own music so that nothing they publish carries copyright risk, and video producers use it to clean up stutters and errors or to supply fully AI-generated voices. We view multimodal AI, and voice in particular, as important, and ElevenLabs' value depends on staying at the cutting edge of human voice communication. The larger opportunity is straight-through processing for content: an AI writes a research paper and its script, sends it through the API to an ElevenLabs voice clone, and passes the audio to HeyGen to produce the video, with each step stitched together automatically.
Max Hodak's Science Corp, a brain-computer interface company and an emerging Neuralink competitor, is showing promise, but we are not yet convinced by the business model for implanted devices. Restoring sight to the blind and treating other neurological diseases and disabilities is valuable work that we want to see succeed, but an investor targeting 3x to 10x returns needs a very large addressable market, and we question whether the patient population is large enough to support exponential revenue growth. A pharmaceutical company can charge a high price for a cancer drug and spread its business across a portfolio of other drugs, while an implant business does not have that cushion. Adoption is the second hurdle: we would not choose to put a chip in our brain today, but we think we might have to at some point, much as late adopters of cell phones and smartphones eventually gave in once everyone else had them, because competitors working at AI speed would be impossible to match. We also doubt that surgical robots can implant chips in billions of people. The case changes with non-invasive devices worn behind or inside the ear, or on the wrist like Meta's neural bracelet, that read neurological signals without surgery. We think a device like that could sell in the billions like the iPhone, which would create a huge addressable market and make companies such as Science Corp and Neuralink very interesting.