Pre-IPO Stock Secondary Market Update | as of Sep 22, 2026 Watch full video = agdillon.com/reports ** Harvey, Glean, Together AI, Baseten, Positron pre-IPO funds accepting investors, closing Oct 2 ... email aaron.dillon@agdillon.com to receive fact sheets (see disclosure at end of email) ** Unconventional AI is attempting something few infrastructure companies dare to try: replacing the compute architecture that has powered computers for the past 60 to 70 years. Founder Naveen Rao, whose MosaicML now accounts for roughly a quarter of Databricks revenue following its acquisition, raised a reported $475 million seed at a $4.5 billion valuation to build brain-inspired, dynamical-system chips designed alongside biologists. The thesis is simple. Energy is the binding constraint on AI, and every token and API call ultimately prices back to power. The human brain runs on roughly 20 watts, and the company's stated target is a 1,000x reduction in inference power. Its first model, Un-0, already matches state-of-the-art image diffusion while running on a software simulation of the chip. Demand is not the problem, with Google alone processing more than 3.2 quadrillion tokens per month and premium AI subscription tiers still selling out. We view energy efficiency as the defining infrastructure trade of the next decade, and the milestone that matters here is working silicon, not simulation. TypeSafe AI attacks the same cost curve from the model side. Founder Diogo Almeida, a former OpenAI researcher behind the InstructGPT and RLHF (reinforcement learning from human feedback) work that became ChatGPT, emerged from two years in stealth with a $40 million DCVC-led seed. Its model, Jev, outputs structured decisions rather than English text, letting machines communicate in their own language instead of translating everything into prose. Pricing is $0.042 per million input tokens with 70 to 500 millisecond response times, and within days Vercel called it the fastest-adopted model in the history of its AI Gateway. The dynamic mirrors China, where GPU constraints pushed labs toward software efficiency and ultimately into open source leadership. Power and token scarcity will now force the same innovation in the US. Both TypeSafe and Unconventional remain early, and our preference is to underwrite companies once product, revenue, and a visible addressable market are in hand. Still, the arrival of two architectures this different is the clearest signal yet that AI remains in its first inning. ElevenLabs Reception marks the voice leader's move from infrastructure into outcomes. The product is a 24/7 AI receptionist that answers calls, books appointments, and speaks more than 70 languages for roughly $29 per month. The economics are striking. One business owner pays $80,000 per year for a single offshore employee whose only job is to call new leads within five minutes, day or night, because speed to lead is his single biggest driver of new revenue. The broader point is that AI applications are converging on outcomes rather than categories. A CEO buying a booked meeting does not care whether it came from a voice company like ElevenLabs, a customer service platform like Sierra, or an enterprise search company like Glean. The winners will be those with the deepest access to company data, workflows, and context, and we expect six to nine months of learning before these agents perform at full capacity. Sales and marketing may prove to be the biggest near-term beneficiary of the AI cycle. Shield AI is reportedly in talks to raise at a valuation of at least $20 billion, up from $12.7 billion in March. The company is guiding to 80% revenue growth and more than $540 million in 2026 revenue, up from $276 million in 2024. Hivemind, its autonomy software embedded across US military aircraft and drones, anchors the software narrative. V-BAT and the upcoming X-BAT, however, make this a hardware business priced at roughly 37x 2026 revenue, a multiple we find difficult to justify outside of a sector leader like Anduril, now at roughly $120 billion implied valuation. Defense budgets are finite, incumbent primes are entrenched, and no startup is likely to capture Google-like market share. The path to sustained growth likely runs through a Palantir-style expansion from government into commercial markets such as stadium and office security. Absent that, we frame defense tech as a durable value franchise with margin expansion and eventual dividend potential rather than a growth story. Castelion illustrates the entry-point discipline required, since a 3x return from a $13 billion valuation implies a $39 billion missile company within five years. * NOTE: AG Dillon ("AGD") is not affiliated with any pre-IPO company in its funds. Some pre-IPO companies require company approval for purchases (aka transfers). AGD has not been pre-approved by any pre-IPO company to purchase their stock. AGD purchases pre-IPO stocks in the secondary market and may gain exposure by directly purchasing the stock (on the company's capitalization table) and/or through a third party fund (aka special purpose vehicle, or SPV). When AGD purchases through a SPV the SPV general partner is responsible for obtaining the pre-IPO stock company approval, not AGD.

Pre-IPO Stock Secondary Market Valuations | as of Sep 21, 2026 | Download full report = agdillon.com/reports | Harvey, Glean, Together AI, Baseten, Positron pre-IPO funds accepting investors, closing Oct 2 ... email aaron.dillon@agdillon.com to receive fact sheets* * NOTE: AG Dillon ("AGD") is not affiliated with any pre-IPO company in its funds. Some pre-IPO companies require company approval for purchases (aka transfers). AGD has not been pre-approved by any pre-IPO company to purchase their stock. AGD purchases pre-IPO stocks in the secondary market and may gain exposure by directly purchasing the stock (on the company's capitalization table) and/or through a third party fund (aka special purpose vehicle, or SPV). When AGD purchases through a SPV the SPV general partner is responsible for obtaining the pre-IPO stock company approval, not AGD.

Pre-IPO Stock Secondary Market Performance | as of Sep 21, 2026 | Download full report = agdillon.com/reports | Harvey, Glean, Together AI, Baseten, Positron pre-IPO funds accepting investors, closing Oct 2 ... email aaron.dillon@agdillon.com to receive fact sheets* * NOTE: AG Dillon ("AGD") is not affiliated with any pre-IPO company in its funds. Some pre-IPO companies require company approval for purchases (aka transfers). AGD has not been pre-approved by any pre-IPO company to purchase their stock. AGD purchases pre-IPO stocks in the secondary market and may gain exposure by directly purchasing the stock (on the company's capitalization table) and/or through a third party fund (aka special purpose vehicle, or SPV). When AGD purchases through a SPV the SPV general partner is responsible for obtaining the pre-IPO stock company approval, not AGD.

Pre-IPO Stock Secondary Market Update | as of Sep 15, 2026 | OpenAI vs Apple Hardware, Positron’s AI Chips & MetaMuse Watch the full video = piped.video/watch?v=kCTnX-xu… The AI inference chip market just produced two very different price tags for the same thesis, and the spread is the most instructive data point in private semis right now. Positron AI raised $875 million at a $5 billion post-money valuation on September 10, roughly five times its approximately $1 billion February mark, and it did so with product in the field: more than 50 first-generation Atlas racks deployed at Oracle Cloud Infrastructure, production customers including Parasail, Jump Trading and i3D, a next-generation Asimov chip taping out on TSMC N3P by year-end, and roughly $100 million of revenue forecast for this year. Etched raised $700 million at a $21 billion valuation in August, doubling its $10.3 billion July Series C in under a month, with Jane Street acting as both lead investor and first paying customer and more than $1 billion in contracts booked, but with no chip yet shipping commercially. Context matters here: Nvidia recently paid a reported $20 billion to license Groq's inference technology and bring founder Jonathan Ross in-house, and the Groq founding team owned less than 10% of the business by the time that outcome landed. That dilution math is the most plausible explanation for Etched's premium, since capital-hungry hardware businesses need enormous funding and founders price early rounds to protect ownership through the build. Our house view is that revenue-generating inference silicon at $5 billion is the more disciplined entry point than pre-scale silicon at $21 billion, and on a 3x-in-five-years framework Positron is well positioned to reach a $15 billion valuation as the Titan system ramps through 2027. We would still encourage diversification across several inference-focused semiconductor names rather than a single position, and we would watch Etched's first production deliveries closely before revisiting the multiple. Meta's September 8 launch of Muse is the first personal AI agent shipped by a company large enough to make adoption a default rather than an experiment, and the market responded accordingly, with the stock up roughly 9% since the announcement and 18% on the month. The product carries a free tier alongside $20 and $100 per month plans, and Zuckerberg has been public about using it well beyond inbox triage, feeding home gym camera footage into the agent for training feedback and using it to hold dietary restrictions and family logistics in working memory when planning travel. The strategic insight is that Meta does not need the frontier model, it needs the frontier application, and its historical strength has been taking technology built elsewhere and getting billions of people to use it. Expect the rollout to proceed through narrow, obvious widgets, a grocery and meal planning widget, a scheduling widget, a kids' activity coordination widget, with the agent eventually proposing them proactively rather than waiting to be configured. The monetization path is the part investors should focus on: Zuckerberg has signaled the product can go free because payments flow through the agent and Meta takes a merchant cut, a Klarna-style take rate rather than a subscription business. The practical constraint on adoption is data governance, and in regulated industries the segregated, air-gapped deployment model remains the right posture for now. For everyone else, the moment ambient agents handle grocery reordering and pantry inventory is the moment this category stops being early adopter technology. The consumer hardware layer is consolidating faster than most models assume, and the acquisition and feature news of the past two weeks makes the direction unmistakable. OpenAI acquired Glass Imaging in a deal reported above $300 million, bringing in the ex-Apple team behind Portrait Mode and computational imaging technology that slots directly into an AI-first consumer device that is always listening, always seeing, and always resident. Apple answered with Audio Intelligence on the Apple Watch Series 12 and Ultra 4, including Siri Recap and Live Rewind, which keep an always-on microphone and retain conversation for seven days with an option to save. That is a genuine posture change from the company that built its brand on privacy, and once the watch establishes the precedent, the same terms of service will follow onto the phone with a single tap of consent. Meta is pushing the same frontier from a different angle, donating Ray-Ban Meta glasses to blind veterans, with users asking the assistant which shirt in the closet is blue and getting a useful answer, which is exactly the kind of demonstration that converts skeptics. The next six to twelve months should bring competing assistants from Apple under new leadership, Google, and OpenAI, with additional entrants from the robotics and devices community. The second-order consequence is the one that matters most for portfolios: as ordinary users experience daily utility, public resistance to data center and power infrastructure buildout weakens materially, and the demand curve for compute and electricity steepens from there.

Pre-IPO Stock Secondary Market Valuations | as of Sep 14, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Performance | as of Sep 14, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Update | Sep 10, 2026 | Uber’s Founder Is Back + Thinking Machines Valuation, Moonshot IPO, and Glean's Efficiency Dream Watch full video = piped.video/5x35BwUM8as Travis Kalanick has emerged from eight years of stealth with Atoms. Atoms is a physical AI company built on City Storage Systems, the parent of CloudKitchens, and the mandate is gainfully employed robots across food, mining, and transport. The founder profile matters here: Kalanick took Uber from nothing to roughly $70 billion before being pushed out, and he brings multi-billion-dollar personal capital and deep Silicon Valley access to a category where funding, not engineering ambition, is often the binding constraint. The supporting data points are accumulating. Figure AI has crossed the threshold where working robots outnumber human employees across the company, and SoftBank is in talks to take a majority stake in 1X Technologies at a reported $6 billion valuation. We would frame the near-term opportunity around specialized robots rather than humanoids, and we would note that the first autonomous robot most people actually interact with will be a car, not a humanoid. Retail appetite has not arrived for robotic pre-IPO stocks. Moonshot AI's move toward a Hong Kong listing will give public investors their first clean read on open-weight model economics. The company behind the Kimi open-source AI models has reportedly filed confidentially and is targeting roughly $50 billion IPO, up from a reported $35 billion at its last raise, on more than $300 million of annual recurring revenue driven primarily by API sales. The competitive position is well understood: Kimi models track three to six months behind the US frontier labs at materially lower cost. The revenue model is the open question. Open-weight providers do not collect a couple hundred dollars a month per seat, individuals run the models for free, and commercial users typically serve them through third-party inference providers such as Baseten or Together AI rather than paying the lab directly. The venue choice is the second question. A Nasdaq listing would have opened the door to Nasdaq 100 inclusion and the passive flows that follow, and $50 billion looks modest against current US frontier lab marks, so the Hong Kong decision likely caps the multiple this business can earn. We would still expect the stock to trade well, because there is almost no pure-play frontier model exposure available in public markets today. Thinking Machines Lab is reportedly raising at a $40 billion pre-money valuation, below the $50 billion plus it sought late last year, on revenue north of $100 million annualized. That is roughly 400 times revenue for a company founded in early 2025 that has already raised $2 billion at a $12 billion valuation in the largest seed round on record. On a risk-adjusted basis we do not like the setup and would not add exposure at this price. The more useful way to read the number is backward. A founder who needs something close to $20 billion of capital to build a frontier lab properly, and who will not accept single-digit residual ownership, has to price the round near $40 billion for the percentages to work. That is venture math rather than a market-clearing valuation, and it explains a growing share of AI mega-rounds. The related dynamic worth tracking is talent flow, because the wealth created inside the frontier labs has reached a level where senior engineers and executives carrying nine-figure balances have no economic reason to take direction from anyone, so departures should be read as founder formation rather than as a distress signal about the companies they leave. The genuinely interesting part of the story is that Thinking Machines is a US-based open-weight provider, a category now filling in around Inkling, Reflection AI, and Nvidia's Nemotron, with Nvidia itself rumored to be evaluating an investment. Glean's vision of an always-on work companion is the clearest articulation we have seen of where enterprise AI value actually accrues. The product thesis is a system that knows your role, your employer, your quarterly objectives, your career ambitions, your weekly tasks, and the people you meet, and that sits inside the messages you send, the documents you write, and the systems you work in, acting proactively rather than on request. OpenAI's newly released GPT-6 Astra points in the same direction, and the six-month timeline now being discussed publicly for ambient AI assistance looks credible. The strategic question for investors is where the margin sits. OpenAI, Anthropic, and Google will all win at the model layer, but context is the scarce asset, and the company that assembles a complete picture of an individual's work life may capture more value per dollar of compute than the labs themselves. The obvious objection is data privacy, and we think it collapses faster than consensus expects. Within twelve months we expect broad opt-in to persistent audio, visual, and screen capture across the major AI platforms, because the productivity benefit will be large enough that users volunteer the access. Glean sits at exactly that layer, and we would weight it accordingly.

Pre-IPO Stock Secondary Market Valuations | as of Sep 7, 2026 | Full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Performance | as of Sep 7, 2026 | Full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Update | as of Sep 4, 2026 | Anthropic's Pre-IPO Spending Spree, Outcome Based Pricing, Instinct Hits $2.5B, Plaud's AI Earbuds Watch full video = piped.video/watch?v=H7SNjfMP… Anthropic is accelerating into the turn weeks before its expected IPO, and the spending tells the story better than any prospectus language will. The company has reportedly agreed to a $45 billion, six-year compute deal with UK neocloud Nscale, covering roughly 460 MW at a West Virginia data center running Nvidia Vera Rubin chips. It is reportedly in late-stage talks to acquire Israeli inference-optimization startup Decart at roughly $7 billion, a business whose entire value is squeezing more throughput out of silicon already installed, and in early discussions with UK chip startup Fractile for inference silicon. The context matters, because Anthropic had the better model and the better product through much of last year, then ran short of compute and ceded share to an OpenAI that had been ridiculed for overbuilding. That lesson is now priced into behavior, since running dry in your first year as a public company is a stock price event rather than an engineering inconvenience. The reported $30 trillion total addressable market framing is aggressive on any conventional basis, though it is less absurd once you accept that AI agents already generate more internet traffic than humans and that robotics adds a second labor pool behind that. Anthropic trades at $1.4 to $1.5 trillion in the secondary market today, and we would not be surprised to see the IPO open above $2.2 trillion, which would clear SpaceX at roughly $1.7 trillion and make it the largest listing on record. Our view is that over-building is now the correct error to make, and we expect Anthropic to keep locking up chips, power, and capacity on every front available to it. Salesforce and OpenAI are both moving away from per-seat SaaS toward outcome-based pricing, where the customer pays only when the software delivers a defined result, such as a qualified meeting booked, rather than paying for seats, tokens, or API calls. This is total alignment between vendor and customer, and it eliminates the single biggest objection we hear from buyers who suspect they are funding an experiment. It also transfers risk onto the vendor, which is itself a signal, because an application company only offers this structure when it is confident its models can stand on their own. We would expect realized pricing to rise rather than fall under this model, since buyers will pay a premium to protect their downside, and that premium is rational on both sides of the table. The second-order effect is a capital cycle: vendors earning on outcomes have a direct incentive to invest more in model quality, which lifts outcomes, which lifts revenue. We think this is the adoption accelerator that finally pulls fence-sitters into AI, and we expect the per-seat model to look dated within a few product cycles. Instinct AI re-rated roughly 5x in a matter of weeks, and the mechanism behind that move deserves more attention than the number itself. The personal AI agent, still in private beta, reportedly raised a $250 million Series B co-led by Index Ventures and Benchmark at a $2.5 billion valuation, up from roughly $500 million in early August. The product reads your calendar, your Notion, and your CRM, then executes, so a request to plan a trip to Paris comes back as sourced flights, a booked itinerary, and a checkout screen. None of this is technically novel, since hobbyist agent setups have done versions of it for a year, but Instinct packages it as an easy button, and packaging is what creates a category. The constraint is not capability, it is consent, because the agent only works if you hand over everything, and early users report it sending emails they did not want sent. We would frame that the way you would frame a new human assistant, which makes mistakes early, accumulates context over time, and becomes indispensable once it knows enough of your history. The closer analogue is social media in 2005, when putting your life online looked reckless and now the absence of a profile reads as suspicious. Our view is that personal AI agents follow the same adoption curve, and that the privacy objection erodes considerably faster than the market currently assumes. Plaud One is a $249 pair of AI earbuds with a 4G-connected recording case, and it reportedly sold out its US pre-sale in a day. Treat it as the same trade as Instinct rather than as a separate consumer hardware story, because an always-on agent is only as valuable as the data it can capture. Capture at a desk is a solved problem, since laptops, phones, and home cameras handle it, but the world outside the office is where the gap sits, and the conference room at a client site is precisely the data an agent cannot currently see. That gap explains why the hardware wave is arriving all at once, with Plaud on earbuds alongside smart glasses, watches, lapel pendants, a forthcoming OpenAI device, and Hark from Figure AI founder Brett Adcock. We expect form factor to fragment along personal preference rather than consolidate around a single winner, which argues for exposure to the data and agent layers rather than to any one device maker. The enterprise side is already normalizing the behavior, with large employers recording employee communications and training models on them, which makes the consumer step feel incremental rather than invasive. The takeaway is that durable value in this wave accrues to whoever owns continuous context on the user, not to whoever sells the microphone.

Pre-IPO Stock Secondary Market Valuations | as of Aug 31, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Performance | as of Aug 31, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Update | as of Aug 28, 2026 | Flock = Safety vs Data Privacy; Reed Jobs Aims to Treat Cancer; Nvidia to acquire Hugging Face; Oura Ring IPO Watch full video = piped.video/watch?v=d3ArV-QH… The public fight over Flock Safety is being framed as a debate about whether data gets captured, and that framing is already obsolete. Flock reports its technology helped locate more than 10,000 missing people in 2025, a figure that spans dementia patients who wandered from home to children in the back of stolen vehicles, and these systems are being voted in by elected city councils rather than imposed from above. The legitimate objection is not capture but misuse, and the reported instances of individual officers querying the network to track former partners are exactly the failure mode that matters. If we can be intellectually honest ... a significant portion of our personal data is freely given to multiple private companies in return for free services (e.g. social media platforms, gmail, corporate employee agreements to track laptop/cell phone activity, etc). The correct mental model is the existing relationship between device makers and law enforcement, where the data exists but a subpoena stands between the state and the record. We would extend the same logic forward: within a few years the always-on assumption will be so widely internalized, first through workplace tooling and then through personal AI assistants, that consent becomes a formality. For investors, that means the durable value in this category sits in access controls, audit trails, and defensible governance rather than in camera count. Oura is reportedly preparing an IPO at roughly a $16 billion valuation alongside a raise in the $3 billion range, which would put nearly 19% of the company on the market in a single print. The revenue trajectory supports a premium: approximately $1 billion in 2025 revenue, up 100% from $500 million the prior year, against a Whoop mark near $6 billion. The instinctive bear case is that Apple simply integrates the functionality into the Apple Watch and compresses the category, and that risk is real, but it misreads what is being bought. Oura sells a subscription against a proprietary longitudinal health dataset, with a genuine niche in women's health around cycle and menopause tracking that the generalist wearables have not matched. The unlock we would underwrite toward is proactive rather than descriptive: wearable data piped continuously to a physician with an AI layer monitoring for exceptions, which is the difference between a dashboard and a product that calls 911. Consider the case of a woman in her twenties who suffered a brain aneurysm and lay unattended for twelve hours; that gap is the product opportunity, and it is not in the current roadmap. We would size this as a software and data business rather than hardware, and we would expect the multiple to re-rate materially when the first credible proactive intervention feature ships. Hugging Face is reportedly in acquisition talks at approximately $13 billion, up from a $4.5 billion mark in 2023, a 2.9x step in roughly three years. The business is poorly understood because it is technical rather than consumer-facing: model hosting, proprietary datasets, training infrastructure, and the default distribution layer for open-source AI. [We now know Nvidia is the acquirer.]

Pre-IPO Stock Secondary Market Valuations | as of Aug 24, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Performance | as of Aug 24, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Valuations | as of Aug 17, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Performance | as of Aug 17, 2026 | Download full report = agdillon.com/reports

Pre-IPO Stock Secondary Market Update | as of Aug 21, 2026 | Etched Hits $21B, Anthropic's Run Rate Explodes to $65B, Higgsfield Raise, & Harvey AI New Product Features Watch full video = piped.video/TZV9ysEpMe8?si=_ikr… Etched's reported jump from a $10 billion to a $21 billion valuation in roughly a month is the clearest signal yet that private capital is repricing inference compute as a distinct and urgent asset class. The company remains pre-revenue, but early reports out of TSMC suggest the technical risk is retiring faster than skeptics expected: the first batch of silicon came back with strong yields, and Etched reportedly had models running inference within four months of receiving chips. Groq, by comparison, sold to Nvidia for $17 billion with an established customer base and a functioning revenue engine, which frames the magnitude of the premium the market is now placing on Etched's forward potential. The talent signal reinforces the conviction, as Etched has reportedly attracted engineers from Nvidia who turned down counter-offers to make the move. Entry-point discipline, however, matters at these levels. At $21 billion pre-revenue, the forward math demands flawless execution across fabrication, delivery, and customer conversion in a category where no full commercial cycle has yet been completed. Positron offers a more conservative entry into the same thesis at a sub-$10 billion valuation with revenue already in hand. If the house view that inference will approach 80% of all compute workloads within a few years proves correct, the prudent approach is diversified exposure across inference silicon rather than a concentrated bet at the top of the range. Anthropic's reported run-rate revenue of $65 billion may represent the fastest revenue acceleration in enterprise software history. The trajectory moved from $10 billion in December to $47 billion in May to $65 billion by mid-August, implying roughly $18 billion of incremental revenue added in approximately two and a half months. Secondary-market prints near $1.5 trillion and growing IPO speculation around a $2 trillion mark reflect where buyers believe this growth curve is heading. The one structural risk worth monitoring is whether cheap open-weight models, particularly from Chinese labs, compress frontier-lab pricing power enough to bend the revenue trajectory. That dynamic, notably, would be bullish for neoclouds and inference infrastructure providers even as it pressures model-layer margins. The broader thesis rests on the expanding size of the AI market itself: bot traffic reportedly now exceeds human traffic on the internet, robotics are coming online with always-on AI demand, and enterprise adoption is compounding as AI-first companies begin to displace slower incumbents. In a market growing at this pace, Anthropic and its peers may cede share to open-source alternatives on a percentage basis while still posting accelerating absolute revenue growth. The path from $10 billion to $65 billion in eight months suggests the S-curve inflection has not yet arrived. Higgsfield's reported $700 million in annual recurring revenue represents a 35x year-over-year growth rate, and the company has quadrupled its valuation to $5.4 billion in just eight months while accumulating 30 million users. The numbers confirm that AI video generation has crossed from novelty into production-grade tooling. Enterprise applications are the near-term revenue driver: a commercial that once required actors, directors, and weeks of production can now be iterated in multiple versions, customized across demographics, and localized with AI-generated lip-sync in any language. The more disruptive force, however, is what this does to the economics of independent content. A feature-length film that would have required millions in production budget can now be prototyped for roughly $5,000 and completed for under $1 million, and early examples are already demonstrating outsized returns on minimal investment. The collapse in content-creation costs is beginning to mint a new generation of independent studios and creators who previously lacked the capital to tell their stories. This is a textbook Jevons paradox: as the cost per unit of content falls toward zero, total content production and consumption should expand dramatically. The firms providing the generation infrastructure underneath this explosion are positioned to capture durable revenue as production volume scales. Harvey's token usage jumped from 1 trillion in January to 14.5 trillion in June, a roughly 14.5x increase in five months that defied consensus assumptions about the ceiling for legal AI adoption. Annual recurring revenue has reached $350 million, the valuation is up 41% in roughly five months, and the company is now rolling out a memory layer that starts with individual lawyers and expands to the broader firm organization. The token-usage explosion is a direct result of forward-deployed engineers compounding automation gains inside law firms: each workflow automated stays automated in perpetuity, and each subsequent workflow adds to a base that never reverts to manual execution. Memory changes the equation further by converting every interaction, preference, and decision into retrievable context, turning what was a capable but amnesiac tool into something closer to a tenured associate who remembers every detail of every engagement. The implications extend well beyond legal. Personal AI devices from Apple, Meta, and Brett Adcock's Hark are coming online as multimodal, always-on systems that will see, hear, and remember everything. The firms that solve memory storage, organization, and retrieval will sit underneath all of it, and enterprises are likely to adopt first by recording and retaining every call, desktop session, and office interaction as proprietary data. Context and memory are emerging as the next infrastructure battleground, and the companies that own this layer are positioned to capture a disproportionate share of the value chain.

Pre-IPO Stock Secondary Market Update | as of Aug 12, 2026 | Robinhood Y Combinator interval fund; Harvey $15.5B raise; Valar raise at $6B; Unitree Robotics $9B IPO Watch full video = piped.video/watch?v=Yo6ssUaZ… Robinhood launched an interval fund investing in Y Combinator backed startups. Retail investors are dying for pre-IPO stock exposure. Companies now stay private through the entire small cap to mid cap to large cap journey that public investors once captured, which means a growing share of total lifetime alpha is created before the IPO ever prices. The delivery mechanism deserves closer scrutiny than the headline. Interval funds are a continuously offered closed-end structure that accepts subscriptions on demand but permits redemptions only quarterly, typically capped at 5% to 10% of total fund assets. In a drawdown that cap becomes a queue, and investors learn the vehicle is easy to enter and difficult to exit. The traditional closed-end alternative is no cleaner: one recently marketed vehicle holding roughly 12% Shield AI and 11% Anthropic traded near a 40% premium to net asset value, a spread driven by retail demand outrunning share supply rather than by anything in the underlying portfolio. Our view is that illiquid private exposure belongs in most portfolios, but sized deliberately and purchased alongside an advisor who can explain premium and discount mechanics before capital is committed rather than after. Harvey's reported raise at a $15.5 billion valuation, a 40% premium to its round just five months ago, is a referendum on the vertical AI application layer rather than on legal software. Revenue tells the story: roughly $350 million annualized, up more than 80% from $190 million in January. The obvious bear case was total addressable market, since legal services alone looked too narrow to underwrite a valuation of this size. Management has answered that by signaling expansion into financial services, insurance, accounting, and tax, which reframes the opportunity as the full set of high-margin professional services verticals rather than a single practice area. The durable moat here is the forward deployed engineer model. Law firms are not technology organizations, so an engineer who embeds and builds the firm's entire workflow becomes structurally difficult to displace, and each additional automated workflow compounds into recurring revenue. Margin expansion is the second leg: aggressive token optimization and intelligent model routing toward lower-cost inference drop directly to the bottom line while customer pricing holds. Harvey is the cleanest public expression we have of the thesis that application companies sitting on foundation models, not the models themselves, capture the enterprise economics. Valar Atomics raising roughly $1 billion led by Sequoia at a $6 billion post-money valuation, plus $200 million in additional financing, marks small modular nuclear moving from a policy conversation into a funded buildout. Two distinctions matter when sizing this exposure. The first is fission versus fusion. Fission is proven commercial technology and the binding constraint is regulatory rather than physical, while no operator anywhere has yet produced commercial fusion. The second is execution culture. When Valar sought the supercomputer that governs reactor shutdown on fault detection, the market quoted $5 million on a two and a half year lead time. The company built it internally in six weeks for $400,000. That vertical integration instinct, borrowed directly from the SpaceX playbook, is precisely what a sector frozen for half a century requires. Energy is the foundational layer beneath the entire AI data center buildout, and we expect regulatory friction to continue falling as power demand compounds. This remains a pre-revenue position, so we would frame it at 0.50% to 1.00% of a portfolio for aggressive investors, not as a core holding. Unitree Robotics listing on the Shanghai Stock Exchange at a $9 billion valuation on $252 million of 2025 revenue prices the company at roughly 35.7 times sales, and the retail tranche was oversubscribed by 5,526 times. That subscription figure, more than the valuation itself, is the datapoint worth carrying forward. Gross margin above 60% on hardware is the second surprise, and it is a function of full vertical integration and domestic manufacture of high-end components such as actuators. The strategic question is whether hardware or intelligence captures the value. Our read is that Chinese manufacturers are winning decisively on volume and component quality while US labs are further ahead on the AI brain that determines actual usefulness in manufacturing and residential settings. That mirrors the large language model cycle almost exactly, where the eventual winners were the firms that paired compute with a frontier model rather than optimizing either in isolation. Regulation now complicates the picture, with a recent executive order restricting sales of Chinese humanoid robots into the US market. Manufacturing throughput remains the real bottleneck for everyone, with leading US producers still building fewer than 100 units per day against a thesis that requires billions of units over 10 to 15 years, and we would note that robotics at scale implies an inference compute demand curve materially steeper than anything software has generated to date.

Pre-IPO Stock Secondary Market Valuations | as of Aug 10, 2026 | Download full report = agdillon.com/reports