Send us Fan Mail
Invest in pre-IPO stocks with AG Dillon & Co. Contact aaron.dillon@agdillon.com to learn more. Financial advisors only. www.agdillon.com
Watch full video = https://www.youtube.com/watch?v=hWRnG3bis6o&list=PLAIaQmt98oP6cyHNemdHcl17WLj3CiIQR&index=1
Anthropic came to market with a target of about $1.4 trillion, and we saw that as a fair valuation. The book was reported at about six times oversubscribed, and the stock priced at $83 per share, above its expected range. The demand does not surprise us: Anthropic was bid in the secondary market the entire time it was private, and we expected that demand to carry into the public market. The structure deserves as much attention as the headline number. Like SpaceX, Anthropic has a staggered release schedule, with an extended lockup for insiders, a regular lockup and multiple early releases, and index providers are now adding these large listings quickly, so early releases line up with index buying at each quarterly rebalance. New listings typically see heavy volatility in the first 30 to 60 days, and we would rather wait for a dip than add on day one, much as SpaceX traded slightly below its IPO price before recovering. We expect Anthropic to do very well from here.
Groq 2.0 is now a pure-play neocloud. After Nvidia's licensing deal of roughly $20 billion, the remaining company buys Groq chips from Nvidia, puts them in its own data centers and sells inference compute to third parties, the same category as Lambda, CoreWeave and Nebius. It has a new cap table, separate from the legacy company, which is buying back stock from original shareholders with one payment still to come. Groq is reportedly seeking a Series B at an $8 billion pre-money valuation, which does not surprise us, and Together AI at $8.3 billion post-money remains an attractive valuation in our view. We are high conviction on neoclouds despite the negative press around their debt. Debt is the point of the model: a data center is funded with equity or debt, and when the math ties, debt avoids diluting shareholders. If a neocloud is profitable today, it is doing it wrong. CoreWeave's backlog reached $132.5 billion at Q3, Crusoe reports more than $140 billion in contracted value, and Nebius says customer prepayments covered over 100% of the capex on deals it signed in Q3. That revenue cannot be recognized until the capacity is built. The market may need a couple of bad quarters to reset expectations, but we believe we are in the first inning of AI and that whoever stands up inference capacity fastest will capture the revenue over the next five years.
Jensen Huang has stepped down from Nvidia after 33 years. We do not expect him to return to Nvidia, and we see the situation as very different from Sam Altman's return to OpenAI. If he chooses to build again, we think he would have no trouble raising capital or recruiting from Nvidia, and a new company led by him would likely be one of the most sought-after pre-IPO rounds in the market. If he joins an existing company, SpaceX looks like the most natural fit given its startup momentum and its push into AI compute.
We see SpaceX as a data center company as much as a space company. It just caught its first Starship upper stage at the tower, and Musk is targeting $500 billion in annualized revenue by December 2027, with Starlink contributing but almost all of it coming from selling data center compute. SpaceX says its compute buildout pays back in less than a year, and given its record of execution, we expect it to keep building as fast as it can. Always-on AI agents show where that compute demand is heading. We run Grok Bot around the clock as our CRM and as a research assistant that monitors the market and turns unstructured data into structured data we can act on, and it supports parts of the day-to-day running of our funds without access to customer or investor data. The key is setup: we treat the agent like a 25-year-old employee, with its own email address, its own Drive folder, a shared drive for joint work, and access to data through APIs and MCP connections. If every person ends up with one agent for work and one for home running all day, the compute buildout still has a long way to go, and that is before robotics.