Deep in the Heart of AI: TomorrowX Conference Takeaways

Deep in the Heart of AI: TomorrowX Conference Takeaways

Market Overview

NPM Private Market Tracker
+110%YTD
Average of the 50 largest names in our NPM Price data
S&P 500
+5%YTD
~−5% week to date through Thursday
Nasdaq 100
+5%YTD
Nasdaq Composite ~−5% week to date

This week Anthropic’s IPO dominated the headlines: the IPO priced Wednesday at $83/share, vs. the $70–$78 filing range, raising $81bn and surpassing SpaceX as the largest IPO ever.*1 ANT opened at $111.50 (+34%), touched $115 intraday, and then faded to close at $97.17 (+17%; $1.8tn mkt cap).*1 The IPO, however, failed to revive the broader AI trade. The Nasdaq rose ~2% early Thursday before reversing into a broad afternoon selloff, with the DJII/S&P 500/Nasdaq Composite closing -2.2%/-2.0%/-2.3%.*2 The S&P 500 is now ~8% below last month’s record, while the Dow is almost ~10% off its August peak. Elevated valuations and the prospect of further Fed rate hikes continue to weigh on equities, even as the 10-year Treasury yield fell to 5.05% (vs. 5.15% a week ago). Week-to-date through Thursday, the S&P 500 and Nasdaq Composite are each down ~5%. Our NPM Private Market Tracker, which shows the average price performance of the 50 largest names in our internal NPM Price data, is +110% YTD vs. S&P 500/Nasdaq 100 +5%/+5%. (Bloomberg Nov, 2026; NPM)

Key Takeaways from TomorrowX Conference

We recently returned from the TomorrowX AI conference held in Austin on November 17-18, where speakers included a range of executives and founders from the broader AI ecosystem. Thematically, the presentations demonstrate that the AI debate has moved from adoption questions to whether returns will justify the buildout and how it will be financed. The most investable shifts we heard include: inference is overtaking training as the driver of compute demand, agents are moving from pilots into day-to-day operations, and cybersecurity is becoming a priority spend as attackers adopt the same tools. For 2027, the “wildcards” appear to be physical bottlenecks (memory, power, interconnect), political risks and the cost of capital. Overall, in our view, the macro tone was more cautious than the AI tone.

Key Themes

1. The overarching AI fear is overbuilt infrastructure relative to returns

The AI trade has been volatile of late, and the central bear case is, not that AI fails, but that capital spending outruns the revenue needed to earn a return. Even the bullish speakers framed their views against that risk.

  • Scott Goodwin (Diameter Capital): More than 100 GW of data centers need financing, including about 40 GW in each of the next two years. That implies $3–5tn of capital (~60% of which goes to chips) against a total corporate credit market of about $12tn. He drew parallels to the telecom (2000s) and shale (2010s) booms, but noted the demand pull from AI is much stronger this time.
  • Marc Rowan (Apollo): If AI revenue disappoints, marginal/higher cost providers will suffer the most. He prefers secured and investment-grade exposure, and financing assets such as power whose cash flows can be underwritten today, over bets on compute pricing several years out.
  • Lisa Su (AMD): Warned of bullwhip risk, where shortage-driven orders turn into excess inventory, while describing demand as “generational”, rather than “cyclical” and still ahead of supply.
  • Gavin Baker (Atreides): Framed the question as a telecom bubble (unused assets) vs. a railroad (a long, useful buildout). His view is that current infrastructure is being used and this is not presently a bubble. Power and wafer shortages and higher rates could slow deployment, which would reduce overbuilding and extend the cycle.
  • David Solomon (Goldman Sachs): Called the current volatility a “recalibration” rather than the “end” of the buildout. He is long-term bullish but does not know what happens over the next three to six months.

2. Inference is becoming the next leg of the AI trade

We have written about this trend in recent weeklies, and the conference confirmed our views. As models move into production, demand is shifting from training runs to serving tokens. Nobody we heard doubted that token demand will explode, and as inference scales, price per token is becoming more important than speed.

  • Brad Gerstner (Altimeter): Frontier labs increasingly face a trade-off between serving paying inference customers today and reserving compute for training the next model.
  • Andrew Feldman (Cerebras): New data center capacity is pricing 40–50% above the 12/25 levels. Cerebras is pursuing 50–200 MW inference sites and disaggregated inference (AMD GPUs for prefill, Cerebras for decode), which he claims yields about 5x the tokens from the same data center.
  • Lin Qiao (Fireworks AI): Fireworks is at ~$1.5bn ARR without owning GPUs. She expects agentic workflows and unlocking institutions’ private data to drive explosive demand.
  • Brannin McBee (CoreWeave): Customers still want older chips for particular workloads; CoreWeave is leasing A100s through 2029, and Blackwell pricing was reported +150% over the past year. More than half of new clients are outside the AI sector. CoreWeave sees potential for a compute futures market.
  • Stanley Druckenmiller (Duquesne Family Office) expects LLM providers to shift from selling tokens to selling completed tasks.

3. Cybersecurity is becoming a high-focus area

The AI tools driving productivity are also arming attackers, and several speakers treated security as both a risk and an opportunity.

  • Scott Wu (Cognition AI). Cybersecurity spend could eventually exceed spend on the rest of software.
  • Stanley Druckenmiller (Duquesne) named cybersecurity, along with healthcare, as an AI opportunity.
  • Lip-Bu Tan (Intel) identified cybersecurity as one of the key risks he is watching.

4. Agents are moving from pilots to the operating model

Agent adoption is still early, but speakers described agents beginning to run real workflows, and argued that broader adoption will be a major new driver of compute demand.

  • Michael Dell (Dell Technologies): The move from computation to intelligence and decisions is a massive shift, and agent adoption will drive higher compute needs. We just are “scratching the surface” on agent use.
  • Lisa Su (AMD): Agents and orchestration raise the importance of CPUs, which organize workloads.
  • Scott Wu (Cognition AI) and Guillermo Rauch (Vercel): Wu said a persistent Devin bot built Cognition’s Android app and keeps it in sync with iOS with no human working on it. Rauch said coding agents let more people build and help each person build more; Vercel crossed $700mm in ARR, +180% y/y.
  • David Solomon (Goldman Sachs) has 13,000 engineers and expects fewer, but more senior engineers in five years. He added that “technological unemployment” is not new and will be managed through, just as it was during the industrial revolution.

5. Financing the buildout is creating a new credit market

The scale of capital required is reshaping fixed income portfolios and widening the cost-of-capital gap between hyperscaler-backed and independent developers.

  • Scott Goodwin (Diameter Capital): The capital need is straining portfolio construction because managers can hold only so much data-center and hyperscaler exposure. It has created asset-backed chip financing as a new asset class. IG and hyperscaler-backed debt currently costs 7.5–8.5%, while independent data center companies pay 12%+. He expects that imbalance to drive consolidation and/or IPOs over time.
  • Brannin McBee (CoreWeave): Backs large capex with long term take-or-pay contracts and asset-backed financing; 20–25% customer prepayments are now standard. He noted that ~130k GPUs come off contract in 2027 and ~170k in 2028, a re-contracting test worth watching.
  • Bret Johnsen (SpaceX): He put SpaceX’s build cost at roughly $30–35 per watt versus $50–60 for competitors (his estimate; we estimate the broader industry at the lower end of this range).
  • Andrew Feldman (Cerebras): A good multiyear cloud contract can pay back in about 18 months.

6. Physical bottlenecks: memory, power, interconnect and fabs

Supply, not demand, is the near-term constraint. Memory looks tightest, and power availability and local politics increasingly determine where capacity can be built.

  • Memory: Micron’s Sanjay Mehrotra said 2027 is “75% sold out.” Speakers noted Nvidia’s tight grip on HBM supply forcing the search for alternatives and driving GDDR6 prices +50% this year.
  • Compute and interconnect: Lisa Su said AMD responded to a slowing Moore’s Law by moving to chiplets. Jitendra Mohan (Astera Labs) said copper is reaching its physical limits as clusters grow, making optics increasingly important.
  • Power: Chase Lochmiller (Crusoe) discussed a 100 GW energy buildout through 2030. Zach Dell (Base Power): “Every energy company is now an AI company.”
  • Community politics: Dina Powell McCormick (Meta) said data-center politics are getting harder with NIMBY opposition. Meta builds only in communities that “want them” and emphasizes local benefits such as tax receipts and skilled jobs. She cited a Louisiana town that funded teacher bonuses of $50k/teacher.

7. Space is emerging as a compute frontier and a defense priority

Orbital compute moved from concept to timeline at this conference, though not everyone is convinced on scale.

  • Bret Johnsen (SpaceX): Last weekend’s Starship landing was a major milestone, and Starship offers a roughly 100-fold reduction in launch cost. He sees an obvious path to orbital compute with initial satellites in 2027, scaling in 2028 and at gigawatt-scale capacity in 2029. Given no energy or cooling costs he sees a cost crossover with terrestrial data centers in one to two years. SpaceX also plans satellite-to-smartphone 5G service.
  • Elon Musk (SpaceX, xAI, Tesla): Described a goal of manufacturing 200 GW per year of solar panels for space, alongside Terafab (first test chip around June 2027) and Optimus robots handling household chores within 12–18 months.
  • Skeptics and supporters: Andrew Feldman (Cerebras) doubts orbital compute reaches the scale being discussed for at least a decade. Stanley Druckenmiller (Duquesne) listed space data centers alongside fusion and curing cancer as potential major AI advances.
  • Defense: Evan Rogers (True Anomaly) said the U.S. is behind China and Russia in space-warfare capabilities and cited a proposed $55bn FY27 Space Force budget that could eventually reach $100bn annually.

8. Where value accrues: applications vs. models

The most consistent view was that durable value will accrue to applications, with a notable dissent from those who expect a few frontier labs to dominate.

  • Applications camp: Lisa Su (AMD), David Solomon (Goldman Sachs) and Stanley Druckenmiller (Duquesne) each pointed to the application layer and real use cases as where long-term value accrues. Druckenmiller sees infrastructure as more of a tactical trade.
  • Model maximalist: Brad Gerstner (Altimeter) expects one to three LLM companies to capture much of AI’s value.

9. Macro and policy are the biggest risks to the AI trade

Speakers were more cautious on the economy and policy than on AI itself. Druckenmiller argued that politics and capital markets, not adoption, are the biggest risks to AI, and several policy decisions (export controls, immigration and Nvidia’s leadership transition) could shape the trade.

  • Stanley Druckenmiller (Duquesne): Has less market clarity than at any point in his career and is in “capital preservation” mode.*3 He sees monetary tightening, widening credit spreads, falling commodities and deteriorating breadth as negatives, but enough innovation that it is hard to be short. He is more bearish than in January 2026, partly because of the Iran war, sees a risk of a disappointing second half of 2027, and worries tighter U.S. immigration policy could hurt growth.
  • Export controls: Lisa Su supports export restrictions on the highest-performance U.S. chips and said demand still exceeds supply despite them. Feldman expects about 75% of Cerebras’s 2027 data centers to be outside the U.S.
  • Nvidia leadership: Solomon said the Nvidia board should take time with its succession decision and not rush to judgment.

Private Market Performance by Industry Sector

Year-to-date return by sector (%)
Industrials
+176.0%
AI & Machine Learning
+148.1%
Consumer
+105.3%
Healthcare & Life Sciences
+72.2%
Enterprise Software
+60.8%
FinTech
+27.1%
Commerce & Marketplaces
+7.8%
Mobility & Transportation
+5.8%
Cybersecurity
−0.8%
Media & Entertainment
−18.2%
Web3 & Digital Assets
−29.4%

* Disclaimer
This commentary is not a recommendation, offer, solicitation of an offer, or advice to buy or sell securities by Nasdaq Private Market, LLC, NPM Securities LLC, SecondMarket Financial LLC, or any of its affiliates (collectively, “NPM”). Securities related services are offered through NPM Securities, LLC (“NPMS”), member of FINRA/SIPC. SecondMarket Financial, LLC (“SMF”) is an SEC-registered investment adviser.

The information should not be taken as tax, investment, legal or other advice, nor should it to be relied upon in making a decision to buy or sell. You should obtain relevant and specific professional advice before making any trade decisions. Registered representatives at NPMS do not provide guidance on the benefits of any specific investment or help with determining the fair market value of securities.

All trademarks are the property of their respective owner(s). “Tape D” and “SecondMarket” are registered trademarks of NPM. Unless expressly stated otherwise, companies listed on this site are not sponsoring, endorsing, or otherwise affiliated with, Nasdaq Private Market, LLC, and none of the statements on this website should be attributed to a listed company.