GM. This is Milk Road PRO, ranking the AI mega-IPO wave just as retail gets its first seat at the table.
Cursor hit $2B in ARR in three years. Anthropic crossed $14B in run-rate. Harvey is inside half the Am Law 100. VCs like those kinds of stats: Q1 2026 set a global venture funding record at $297B, with 80% going to AI.
The headline hides the structure. Four deals (OpenAI, Anthropic, xAI, Waymo) absorbed $188B, or 65% of the total. The other 6,000 startups split the rest. Median seed-stage AI valuations sit 18% below a year ago.
The 10 largest AI companies carry over $2T in combined value. None are public. That's about to change: Cerebras prices this month, SpaceX targets a June roadshow, Anthropic and OpenAI are filing for late 2026 to early 2027.
My take: access is finally here. Most of these IPOs will price too rich to buy on day one. The trade worth watching is what gets sold to fund them.
So, here's what Iโm covering today:
- ๐ The four deals that ate Q1.
- ๐๏ธ Where AI capital really goes.
- ๐ How retail gets in.
- ๐ Three vehicles, three trade-offs.
- ๐ The 2026 IPO pipeline.
- โ ๏ธ The risk under the pipeline.
- ๐ฏ What Iโm watching, what Iโm skipping.
Iโm also including a full review of my current Milk Road PRO Portfolio!
Scroll to the very end to check it out, including why I sold a few tech stocks this week.
If you want to see what else Iโm holding and watching, you can pay just $1 for a 7-day trial of Milk Road PRO.
The four deals that ate Q1
The headline number has no modern precedent. Global venture funding hit $297B in Q1 2026, a 2.5x jump from the prior quarter, with $242B of that going to AI. To put it in context: Q1 alone outpaced every full year of global VC activity prior to 2019.
But underneath the headline, the picture flips.
Four deals, 65% of capital
Q1 saw four of the five largest venture rounds ever recorded:
- OpenAI: $122B at an $852B post-money valuation.
- Anthropic: $30B at a $380B post-money valuation.
- xAI: $20B Series E (total raised now $42.7B).
- Waymo: $16B at a $126B post-money valuation.
Together, those four rounds absorbed $188B, or roughly 65% of all global venture funding for the quarter. Strip them out, and the remaining 6,000-or-so startups split the rest.
What the headline misses
The early-stage AI market is tighter than it was a year ago, even underneath the all-time funding record.
Median seed-stage AI valuations in March 2026 sat 18% below March 2025 (S&P Global). Global deal count fell 15% quarter-over-quarter to its lowest level since late 2016. Non-AI venture funding, inflation-adjusted, is below Q1 2020 levels.
The shape looks like the early oil industry: a small handful of infrastructure giants absorbing the capital, while everyone else competes to build on top of their platforms.
When 65% of capital flows to four companies, access becomes the variable that matters most for investors.
Where AI capital really goes
The AI startup universe is big, but the money clusters into three layers, each with very different economics.
Layer 1: Frontier Labs (the foundation)
These are the companies training the giant models everyone else builds on: OpenAI ($852B), Anthropic ($380B), xAI (around $200B post-merger with SpaceX), Mistral ($14B), and Google DeepMind inside Alphabet.
The numbers are eye-watering, but so is the burn. OpenAI is reportedly projecting a $14B loss for 2026, has committed to spending $1.4T on data centers over eight years, and is funded primarily by debt and circular vendor financing (more on that risk later).
Bull case: Foundation models are the new utility layer.
Bear case: they commoditize. DeepSeek rattled the market in early 2025 by training a frontier-quality model for a fraction of U.S. costs, and open-source models from Meta and Mistral keep the gap narrow.
Net takeaway: highest absolute upside, highest absolute risk. Concentrated capital, concentrated outcomes.
Layer 2: Vertical AI (knowledge work agents)
This is where the most interesting business models live.
Vertical AI startups take a foundation model, wrap it in domain expertise, and embed it inside a specific industry workflow. They charge like enterprise SaaS (multi-year contracts, high net revenue retention) and grow like consumer apps.
The standouts in early 2026:
- Databricks: Horizontal data / ML infrastructure.
- Cursor (Anysphere): AI-native code editor.
- Cognition (Devin): Autonomous coding agents.
- Harvey: Legal AI for large law firms.
- Legora: European legal AI platform.
- Abridge: Ambient clinical scribe for healthcare.
- Sierra: AI customer service agents.
- Decagon: AI support automation for enterprise.
- Mercor: AI-driven hiring and talent marketplace.
The pattern: pick a regulated, high-trust industry (law, medicine, finance, support), build a workflow agent that beats the human baseline, charge enterprise prices.
Bull case: real revenue, real moats (data, integrations, switching costs), realistic IPO paths in 2026-2027.
Bear case: foundation model providers can absorb the application layer (Anthropic's Claude Code is already the fastest-growing competitor to Cursor).
Layer 3: Physical AI
We covered this segment in our last reports. Q1 2026 confirmed the thesis. Physical AI now has real venture money and real product behind it.
Q1 highlights for massive raises:
- SpaceX: Reusable rockets and data centers in orbit.
- Waymo: Robotaxi service and self-driving cars.
- Skild AI: General-purpose robotics intelligence model.
- Physical Intelligence: Foundation models for robot manipulation.
- Saronic: Autonomous naval drones.
- Anduril: Autonomous defense systems and software.
- Helsing: European defense AI.
- Wayve: End-to-end self-driving software.
Physical AI led all sectors with 11% of Q1 deal share.
This layer is the longest-duration bet (hardware cycles are slow, capital intensity is brutal), but the eventual market (robotaxis, humanoid labor, autonomous defense, industrial automation) is the largest end-market AI will touch.
How retail finally gets in
The hottest companies in the world have no public market presence. The 10 largest AI startups carry over $2T in combined value. None are public.
OpenAI alone, at $852B, is bigger than every company in the S&P 500 except a tiny handful at the top. Anthropic at $380B is bigger than Coca-Cola, Disney, or Goldman Sachs.
For most of the last decade, retail investors had no way in. You needed accreditation, top-tier VC relationships, and minimum check sizes starting at $100k.
That's starting to change.
Three vehicles, three trade-offs
There are three real routes for getting AI startup exposure as a non-accredited investor.
Route 1: Direct holders of AI startups in public form. The cleanest, lowest-friction path. Public companies that own meaningful slices of the private AI giants:
- Microsoft (MSFT): ~27% stake in OpenAI.
- Alphabet (GOOG): Wholly owns Waymo, plus DeepMind, plus Gemini.
- Amazon (AMZN): Major Anthropic investor (~$8B committed), plus AWS as the inference rails.
- NVIDIA (NVDA): Picks-and-shovels, plus equity stakes in OpenAI, xAI, CoreWeave, and others.
One caveat: Microsoft, Alphabet, and Amazon don't trade on AI alone.
Each has massive non-AI businesses (Office, Search/YouTube, retail/AWS) that drive most of the share price. That's a feature for diversification, but it means AI gains can be diluted by what's happening in the rest of the business. NVIDIA is the closest to a pure-play AI bet among the four.
Route 2: AI-themed ETFs. Diversified bundles. Lower single-name risk, broader thematic exposure:
- AGIX (KraneShares AI & Tech): some Anthropic exposure plus a basket of AI-leveraged public names.
- XOVR (ERShares Private-Public Crossover): hybrid structure with some SpaceX and OpenAI exposure.
Route 3: Pre-IPO access funds. The newest and highest-risk option. Three publicly available funds offer retail-accessible exposure to private AI companies.
Premium-to-NAV dynamics, illiquidity, single-name concentration, and high fees mean these can lose money even when their underlying holdings perform well.
RVI trades at fair value, so there's no premium drag. But only 15-20% of the fund is AI. The rest is fintech. The C-corp structure adds another layer of cost: gains get taxed twice, once at the fund level and once when distributed to you.
DXYZ has the best underlying holdings (52% SpaceX, plus OpenAI), but you pay a 33% premium for them. Every dollar you put in buys roughly 75 cents of actual assets. When SpaceX IPOs, that premium will compress sharply.
ARKVX has the deepest AI exposure (30-40%) and the brand-name holdings you want (Anthropic, OpenAI, xAI, Figure AI). The cost is the catch. A 4.7% annual fee compounds heavily over multi-year holds. And the interval structure means you can only redeem during quarterly windows.
The 2026 IPO pipeline
The vehicle comparison above is a snapshot of today. The IPO calendar over the next 12-18 months could rewrite it. This is the deepest mega-IPO pipeline in modern equity capital markets history.
What's filed or filing
Ranking the IPO pipeline
Ranking the pipeline from most to least attractive at the moment of issuance. I think the order will surprise you.
The most boring deal is one of the most interesting ones. Databricks is FCF-positive, growing 65% on a $5.4B run-rate, with none of the capital-structure drama that surrounds the others. If it prices reasonably it's the kind of deal that compounds โquietlyโ for years. A high conviction "buy and hold" candidate of the group, and it almost never gets discussed because the story is so boring.
SpaceX is next, but only at the lower end of pricing. The $1.75T target is rich, but the underlying business is genuinely diversified across Starlink subscriptions, government contracts, xAI, and commercial launch. The 30% retail allocation also eases the supply-demand crunch that usually inflates these mega-IPOs. The major wildcard of course are data-centers in space. The biggest risk is that similar to other Elon Inc. companies (e.g., TSLA) it takes more time until he fulfills his promises. Also, if you are a TSLA shareholder, there is a very reasonable chance you get exposure to SpaceX anyway via an upcoming merger. Thatโs my base-case, which is why Iโm not buying.
Cerebras sits between the infrastructure plays and the model labs. I like it generally, because it provides physical infrastructure (chips), which is favored by the market at the moment. The wafer-scale architecture they provide is differentiated from peers like NVDA, and the OpenAI compute deal (up to 750MW over three years, worth more than $10B) validates demand. The problem is two-fold: on the one hand, they are in a very competitive market and on the other, they have a very aggressive targeted sales multiple (~up to 69x vs. NVDAโs 25x).
Anthropic has a clean fundamental story. $14B run-rate growing 10x annually, increasingly priced into enterprise budgets via Claude Skills. Valuation is everything here. At the $380B private mark it's already aggressive, so the question is whether the IPO prices a re-rating in or not. A muted opening would create a buyable entry, if you believe AI models will not get commoditized (I personally do, which is why Iโm not buying into it).
OpenAI is where caution ramps up. The fundamentals are impressive. ~$14B in 2025 revenue, dominant market position, but the financial structure is heavy. $14B projected loss in 2026, $1.4T in data center commitments, and the circular financing chain laid out below. Worth considering only at a valuation that meaningfully prices in the burn, with a multi-year hold horizon rather than a flip.
The risk under the pipeline
Every valuation in the ranking above assumes the AI build-out keeps compounding. That assumption deserves scrutiny.
The โAI flywheelโ is really a closed loop. Most of the dollars keep moving between the same handful of companies. NVIDIA invests in OpenAI. OpenAI buys NVIDIA GPUs. Microsoft owns 27% of OpenAI and is also ~20% of NVIDIA's revenue. And so onโฆ
Optimists call it vendor financing that lines up suppliers, builders, and customers. Bears call it the structure that defined the late-stage dot-com bubble.
The impact/demand picture is mixed at the moment.
Impact-wise, a February 2026 study found 90% of firms reported no measurable AI impact yet. Bain estimates AI will need ~$2T in annual revenue by 2030 to justify current infrastructure spending. Sam Altman himself said in August 2025 that investors are "overexcited."
But the demand numbers expressed in cash flows are real.
NVIDIA delivered $68.1B in revenue in fiscal Q4 2026, up 73% year over year, and Anthropic is at a $14B run-rate.
The demand backlog disclosed on Q1 2026 earnings calls is concrete: hyperscaler CapEx guidance (Microsoft, Alphabet, Amazon, Meta combined) topped $410B for the year, with all four citing capacity-constrained AI workloads as the gating factor on growth. Enterprise customers are waiting in line.
So, in my view, the closer historical comparison is the telecom build-out of the 1990s. Brutal bust for early infrastructure investors, but the fiber and the towers were eventually justified by demand.
So, the infrastructure investments may be right. The companies financing it may not all survive.
What I'm watching, what I'm skipping
Access to the shiny private AI companies is possible. That does not make these vehicles attractive.
In my portfolio, I am not buying these vehicles. Fees are too high, transparency is too low, and the extra private-equity risk is not worth it when public markets already offer a wide set of AI exposures with cleaner reporting and very attractive returns.
The same caution applies to the IPO wave. These deals will be flashy, heavily marketed, and priced rich. That does not make them good buys on day one.
Off the IPO pipeline I would be mainly looking at Databricks and SpaceX, because their business models feel more differentiated and less easily commoditized than the frontier labs. Databricks sits close to the enterprise data layer, while SpaceX owns real physical infrastructure through launch and Starlink. Those feel like stronger moats.
The model companies are harder for me. They may be hugely important, but the commoditization risk is real, and the capital intensity is brutal. If staying competitive means constantly taking on more debt and more compute commitments, I do not think public markets will reward that (see hyperscalers today).
The more interesting setup may be the second-order one.
These mega-deals could pull liquidity out of the rest of the market, as (especially retail) investors may sell existing AI winners to fund new positions. That dislocation creates better opportunities outside the IPOs than inside them.
What I'm watching:
- SpaceX pricing toward the lower end of its $1.75T target ($1-1.2T).
- Databricks IPO opening. A muted print is buyable.
- Dips in the wider AI space as liquidity rotates into the IPOs.
Until then, my view holds: access to those private plays is here, but most of these IPOs will price too rich to buy on day one. The trade worth watching is what gets sold to fund them.
Portfolio update: I sold a little bit
Every Milk Road AI PRO report will now close with my real portfolio moves from the last two weeks. The thesis lives in the main report above. The trades live here. You see exactly how I'm acting on what I'm writing.
The IPO report above is my prep work for the wave of mega-deals coming to public markets. The moves below stand on their own. They cover what I've been doing across the rest of the book over the last two weeks.
What I'm selling
Uh, Ohโฆ ๐ง The rest of this report is exclusive to PRO members!
Already a PRO member? Log in here.
Want to see what I sold and why? Milk Road PRO is only $1 right now, so get on in here and scope my full portfolio!
