Opinion

The Reason Prior Tech Bubbles Broke Just Got Fixed.

Every prior tech bubble was funded by venture capital alone, and broke when VC ran out. This one isn't. Big Tech alone is spending $725B on AI in 2026, more than four times what all of venture capital deploys in a year. That single fact ends the bubble argument.

Wall street just took AI from VC, next all of tech

Nobody Told the Bubble Callers. So What Does That Tell You About Belief?

Stephen Messer
Co-founder, Collective[i] & Intelligence.com. Co-founder, LinkShare (sold to Rakuten, $425M). EY Entrepeuner of the Year/Deloitte Fast 50 Winner (2x) Board Member, Spire Global (NYSE: SPIR). Join Intelligence.com

Everyone I know spent this weekend drafting their fantasy football team. My family included. Watching venture capital try to fund the AI era is starting to feel like watching my nephew play fantasy football. Among his peers, he looks like a genius. He knows every stat, every matchup, every sleeper pick. Its fun to play but its not playing at a real NFL game level. Different game. Different stakes. Different players.

That is the exact gap that separates venture capital from the greater finance industry that just showed up for AI and is spreading through all of tech. VC has been the best game in tech for two decades, it was a cottage industry with cottage eccentricities. Among its own peers, the largest VC's still looks like geniuses. Then Big Tech entered their game, committed $725 billion in 2026 capex. Wall Street jumped in and filed compute futures at CME. BlackRock started leading quantum rounds. Sovereign wealth funds started writing $10 billion single checks. The pros showed up. Compared to the balance sheets now on the field, VC is a fantasy league.

Which is why every time I get on a call with an investor and hear the word bubble coming out of their mouth, I know I am talking to somebody playing in the wrong league, watching the wrong players, and not realizing the game is now on a different level. Bubble talk is muscle memory from 1999. In the VC game you dream of massive growth at value investing prices. That existed a long long time ago. Its why tech had so many boom and busts. I lived through 1999 and I still have the scars, which is precisely why I can tell you what is different now.

AI opened a door that all of tech is passing through. Tech will get bigger because of it. Tech will get faster because of it, and our economy will benefit from these change for the most part (there will be negatives inevitably but for now mostly good). The next AI models will scale through ecosystem structures that never existed in prior cycles, because the financial architecture underneath is finally big enough to fund it that way. Every prior tech bubble in modern memory was funded by venture capital alone. This one is not. Big Tech, Wall Street, Big PE, sovereign wealth, and the exchanges have all shown up with capital structures venture capital cannot supply. That single change is enough to end the bubble argument on its own.

This piece is the setup. The AI first world is showing leaders that belief/gut is the new risk in every senior seat, and AI is the amplifier that makes belief lethal at speed. A three-part series is coming right after this one on how the skills required at the top of every serious company have changed. My argument below on why there is no bubble are designed to show just how much is changing and why our belief/guts are causing smart people to show bad judgement, the next series is going to give you the vocabulary that separates you from everyone else. If you finish this piece still thinking bubble talk is what smart people say in a room, there is nothing I can write next that is going to help you.

Every Prior Tech Bubble Was Funded by VC Alone. This revolution is bypassed the VC and grew larger

Here is the thing every bubble caller keeps missing. The dot-com bubble ran on venture capital. The 2014 SaaS wave ran on venture capital. Crypto in 2017 ran on venture capital plus retail. Web3 in 2021 ran on venture capital plus retail plus a light coat of hedge fund exposure. Every one of those bubbles broke the same way. VC ran out of the appetite or the money to keep funding the next round, the marginal buyer disappeared, and the whole thing collapsed inward on the funding structure that built it. The bubble arithmetic requires venture capital to be the only serious buyer. When it is, the crash is a matter of when.

The 2026 AI environment does not match that pattern. Venture capital is still writing checks, but venture capital is now the smallest financial player at the tech table. The four largest hyperscalers alone, Amazon, Microsoft, Alphabet, and Meta, are guiding a combined $725 billion in 2026 capital expenditure, up 77 percent from $410 billion in 2025. Alphabet issued rare 100-year bonds to help fund the buildout. Wall Street investment banks are arranging debt for hyperscaler capex programs at scales that dwarf anything the venture industry has ever underwritten. Sovereign wealth funds are directly investing at the $10 billion single-check level. Big PE has over $2 trillion in dry powder and is starting to sign warrant-swap structures with model companies. The exchanges are preparing to price compute as a commodity. Nvidia itself has become a financing counterparty, with an investment portfolio that hit $95.6 billion at the end of July, up from less than $100 million in early 2020. Every one of those pools is deeper, more patient, and better-hedged than venture capital was in any prior cycle. That single change is enough to break the historical comparison the bears keep making.

I have covered pieces of this shift in prior articles. Tech Is Not an Asset Class Anymore laid out why tech is now the economy itself rather than a slice of it. The Trillion-Dollar Trade Wall Street Isn't Seeing laid out why every data center is a free call option on Taiwan risk. Peak Token laid out why frontier commoditization is a feature of the market. Anthropic hit $65 billion in annualized revenue in July, on its way toward $100 to $120 billion by December. Bubble metrics do not look like that. That is a company earning more revenue in three years than any software business in history has ever produced, and the balance sheets underwriting it are not the ones that funded pets.com.

Every prior tech bubble was funded by venture capital alone. Every one broke when VC ran out. This one is not being funded that way, and pretending otherwise is going to cost the doubters more than they think.

AI, Quantum, and Space Weren't Built by VC. Not Much at This Scale Ever Is.

The venture-capital-only pattern was a story that ran through a specific window of software history. Look at anything bigger than software and it stops holding. AI. Quantum. Space. Semiconductors. Biotech. None of those industries were built primarily on venture capital, because venture capital is not sized to build them.

Start with AI itself. Nvidia alone committed over $40 billion in AI equity investments in the first four months of 2026, with $30 billion of that in a single OpenAI round. In the same window, the entire US venture capital industry deployed roughly $57 billion into every non-AI startup category combined, fintech, biotech, climate, SaaS, consumer, per Crunchbase Q1 2026 data. One chipmaker's AI equity deployment ran at close to the same scale as the entire non-AI venture industry. Nvidia's non-marketable equity holdings alone grew from $3.39 billion at the end of January 2025 to $22.25 billion a year later, a nearly seven-fold increase in twelve months. Anthropic's cap table reads more like a Fortune 100 balance sheet than a startup roster. Microsoft holds the largest single stake in OpenAI. Amazon has committed $8 billion to Anthropic across two rounds. Google is on the same cap table for billions. Sovereign wealth funds are writing $10 billion single checks. Big Tech is guiding a combined $725 billion in 2026 capex, most of it AI infrastructure, with Alphabet issuing 100-year bonds to help fund it. That is not a venture-backed sector. That is a nation-state-scale infrastructure buildout financed by the deepest balance sheets in the world.

Now quantum. PsiQuantum, one of the leading fault-tolerant quantum computing companies, hit a $10.5 billion valuation in May 2026 after a $1.5 billion round, on top of a $1 billion round in September 2025 at $7 billion. Look at who is writing those checks. BlackRock led. Temasek co-led. Baillie Gifford. Qatar Investment Authority. Macquarie Capital. Nvidia's NVentures. Morgan Stanley's Counterpoint Global. The Australian and Queensland governments committed $617 million on their own. Founders Fund and a handful of traditional venture firms sit on the earlier rounds. The growth capital is coming from the largest asset managers, sovereign wealth funds, and corporate venture arms on the planet.

Now space. SpaceX did not become a $350 billion company on venture capital. Google put $1 billion in during 2015. Fidelity, T. Rowe Price, and later sovereign wealth funds wrote most of the growth capital. NASA contracts underwrote the launch business that made the rest possible. Blue Origin has been funded almost entirely by Jeff Bezos personally. Rocket Lab is public. Anduril's space defense business is a mix of defense contracts and strategic capital. Spire Global, which I sit on the board of, went public through a SPAC, and I wrote about how that whole class of company got built in The Shoebox That Changed Space. The venture capital pattern only ever fit software, and only during the specific window when software could be built for $5 million and shipped for $50 million. AI is now joining every other major tech sector in being built by the deep balance sheets. The bubble callers are pattern matching to the wrong industry.

Wall Street Innovation Just Turned Compute Into a Commodity.

The single most important AI story of the summer got buried under earnings coverage. Wall Street is preparing to put a tradable price on compute. CME Group plans to launch futures contracts on October 5, pending CFTC approval, tied to the hourly rental cost of Nvidia's H100 and B200 GPUs. The contracts will settle against benchmarks from Silicon Data, which currently prices H100 rentals around $2.68 per GPU-hour and the newer B200 around $5.66. Jared Blikre at Yahoo Finance walked through the mechanics in a piece worth reading in full.

The reason this matters has almost nothing to do with the specific product. A single commodity futures launch does not reshape a market. The reason it matters is that compute is now being publicly recognized as a commodity in the exact same category as soybeans, cocoa, natural gas, and oil. When a commodity gets a futures market, a specific set of things happens next. Producers hedge production. Buyers hedge input costs. Countries hedge strategic exposure. Speculators provide liquidity. Prices become discoverable. Capital allocation gets easier. Infrastructure gets built faster because financing becomes cheaper. Every one of those effects has played out in every other commodity market in modern history. There is no reason compute is going to be an exception.

The Southwest fuel hedging story is the cleanest analog. Southwest famously locked in fuel prices for years using futures contracts, giving the airline a structural cost advantage during oil price spikes that other carriers could not match. Every major AI company is about to have the same option. Anthropic can hedge its compute exposure through 2029. Amazon can hedge inbound compute demand from its AWS customers. Nvidia can participate in the rental economics of the chips it sells. Countries with strategic AI ambitions can lock in national compute inventories the way they currently lock in oil reserves. The financial market goes from broken to functional in one launch date. Risk is now more evenly spread and thus derisking a bubble like crash.

The precursors are already trading over the counter as bespoke bilateral deals. Anthropic signed a $45 billion contract with SpaceX in May, paying $1.25 billion per month through 2029 for 220,000 Nvidia GPUs at Colossus 1 in Memphis. Google followed a month later with a $920 million per month deal with SpaceX for 110,000 GPUs through June 2029. Those are compute contracts in everything but name. The futures market is going to formalize what the largest players are already writing by hand.

A routing marketplace comes next on top of that curve, doing for GPU capacity what Stripe's $7 billion OpenRouter acquisition did for model access. For the bears betting that compute prices collapse and take Nvidia with them, this is where the trade gets uncomfortable. A functional futures market means natural hedges emerge, and natural hedges kill the crash scenario the shorts are pricing. Every past commodity that got a real futures market saw its price volatility decline over the following decade. Nvidia already sits at the center of that market. As Blikre put it in a follow-up piece, Nvidia is starting to look like the central bank of AI. Jensen Huang described the shift in one sentence. "Now, compute is revenue."

Compute is now being publicly recognized as a commodity in the exact same category as soybeans, cocoa, natural gas, and oil. Every effect that follows for other commodities is going to follow here.

The Chip Demand Curve Has Ten Buyers Behind the Frontier.

The bear case reduces the entire compute build to one buyer. The frontier LLM providers. If OpenAI or Anthropic slows down, the story goes, the whole compute market catches a chill. This misreads what a GPU is.

A GPU is general-purpose. That is a mechanical fact, and it is the reason crypto miners were able to pivot into hyperscalers in 2023 without changing their hardware footprint. The same silicon that runs an LLM runs an economic prediction model at Collective[i], folds a protein at AlphaFold or Isomorphic Labs, controls a robot at Physical Intelligence or Figure, plans a route at Waymo, predicts a hurricane at GraphCast, screens a mammogram, or designs an alloy. I walked through the taxonomy in The Companies Winning at AI Are Playing a Different Game. Roughly a dozen categories of model each generate their own demand curve. Frontier LLMs are one of them. Nvidia and the memory providers are the closest players to that curve, and they keep raising guidance because the demand runs ahead of what a frontier-LLM-only story would produce. Data Center revenue at Nvidia hit $89 billion last quarter, with total revenue guided to $108 billion for the current quarter. That reads as a diversified demand curve with more customers than the market is pricing in.

Then there is the Taiwan risk, which I already covered in The Trillion-Dollar Trade. If one island stops shipping chips for any reason, the demand from the rest of the world for existing GPU inventory becomes enormous. Every current data center becomes a strategic asset priced at a call option on that scenario. The bear case has to be right about frontier LLM demand collapsing at the same moment that robots, self-driving vehicles, drug discovery, materials science, weather modeling, and geopolitical risk all simultaneously fail to fill the gap. That is not a bet I would take.

The Death of SaaS Just Got a Bellwether Announcement.

The dead cat bounce in SaaS is now in full effect, and every serious investor knows it. I called the shape in The Last Great Head Fake in Software History. The odd part about the current moment is that the same belief structure producing bubble talk about AI is also producing the SaaS-as-undervalued play. Two beliefs. One investor. Both wrong. The same person who is sitting out AI because it "feels like 1999" is often the same person loading up on legacy SaaS because it "looks cheap." Belief takes them out twice. This logic leads to a Correlated Breakdown, not a hedge.

Look at the last three years for the version of this that already played out. Not one VC firm I know of hedged its SaaS portfolio while multiples slid from 20x revenue in 2021 to bargain-basement over three years. I am not sure most of them even knew how to hedge that risk. Wall Street traders hedge falling assets every day. Family offices structure downside protection into every allocation. The professional finance industry treats hedging as basic hygiene. VC treated its SaaS book like the only tool available was hope. So the same firms that missed AI on the upside also failed to protect the value they already had on the downside. Three years is a long time to hold a losing position without a hedge. It is exactly the kind of thing an NFL team's front office would never let happen, and exactly the kind of thing my nephew's fantasy league does every week.

On August 26, 2026, Salesforce and Anthropic announced Claudeforce. Marc Benioff and Dario Amodei took the stage together. Salesforce called it "the world's #1 AI meets the world's #1 AI CRM." Claude becomes the default reasoning model across Agentforce, Slack AI, Slackbot, and Agentforce Coworker. Salesforce becomes a plugin inside Claude with 37 prebuilt sales skills. Salesforce is spending roughly $300 million a year on Anthropic tokens and holds a stake in Anthropic reportedly valued around $5 billion. What was crazy was the the street liked the announcement driving SFDC stock higher, the dead cat bounce.

In practice, this may be the dumbest strategic move Salesforce has made in twenty years, and may be remembered as the beginning of their end. I say that as someone who has watched Marc build one of the great software companies of my generation. Salesforce had a killer quarter. The company then walked to the podium and announced it was handing the brains and the interface of its product to the company that is going to eat it. The one thing Salesforce had left, the thing that justified its enterprise pricing for a decade, was lock-in. Data lived in Salesforce, workflows lived in Salesforce, the seller worked inside Salesforce. Handing the reasoning layer and the seller-facing interface to Claude means the seller now works inside Claude, with Salesforce as a plugin serving data upward. The direction of the arrow is what matters. The customer relationship, over time, belongs to whoever owns the interface.

Now contrast Salesforce with the one software incumbent that saw the AI wave early, has the operating experience to scale around it, and has the balance sheet to hedge across every layer. Microsoft. Microsoft bought the largest single stake in OpenAI. Microsoft built OpenAI into every product it ships. Microsoft is now visibly diversifying at every layer above and around that partnership. GitHub Copilot added Kimi K2.7 Code from Moonshot on July 1, the first open-weight model in the Copilot picker. Microsoft is evaluating Kimi K3 for Azure Copilot after its July release. The Azure Foundry catalog now sits DeepSeek, xAI, Meta, and Moonshot alongside OpenAI and Anthropic. Microsoft has not walked away from OpenAI publicly. Microsoft has stopped needing to bet the company on them. That is the more consequential move.

If you were banking your capital on where the world is heading, would you bank on Benioff, who just tied his flagship product to a single closed-model provider at the moment the pricing floor is about to fall out? Or on Nadella, who is hedged across half a dozen providers and pays the marginal cost he wants to pay on any given workload? That question deserves to be asked every quarter for the next four to six. Out loud. In every board meeting where either name is on the strategy slide.

If that point did not leave its mark, I want you to look at what China is already starting to do to the closed-model pricing floor Salesforce just bet on. China has run the same dumping playbook in steel, aluminum, solar panels, and electric vehicles so I would expect something similar here. Flood the market with subsidized capacity. Undercut Western producers on price. Capture share. Sustain losses through state support until Western incumbents cannot follow. The open-weight AI push is the same playbook applied to models, and it pairs with the most important policy goal outlined in China's last policy plan. Fund a nascent Chinese chip industry by giving it a global demand base. The open-weight model is the demand generator. The Chinese silicon underneath it is the product actually being sold. Salesforce just built its entire agentic strategy on the assumption that closed frontier APIs stay dominant, at the moment the Chinese Government made dumping cheap AI its priority. That assumption has a shelf life measured in quarters. I called the shape of the SaaS collapse in Software Is Not Going Down Alone. Claudeforce is the bellwether. The dead cat bounce has a face now.

Belief takes them out twice. The same person who is sitting out AI because it feels like 1999 is often the same person loading up on legacy SaaS because it looks cheap.

Belief Is the Common Thread.

Add the above mistakes together. Bubble talk is a belief that this is 1999. SaaS as an undervalued play is a belief that yesterday's revenue multiples price tomorrow. Salesforce betting on closed frontier APIs is a belief that today's pricing floor holds even as the gap in frontier intelligence is all but gone. All three come from the same investor mindset. Yesterday's playbook, run harder, produces tomorrow's returns.

Notice what I am not saying. I actually think the frontier LLM business itself is going to commoditize hard, and I made that case in detail in Peak Token. Pure-play API-only frontier labs face real downside. Anyone reading the archive knows I have argued this in detail. And with all that, I still do not think this is a bubble. Those two views coexist because the money funding this cycle can absorb LLM compression and still fund what comes next.

Traditional VC was always long-only, all-or-nothing, and structurally unable to hedge, take debt, or write differentiated exposure. Every prior tech cycle rose and fell as a single binary because the financial instrument funding it was a single binary. Real finance does not run on binaries. It runs on structured exposure to a range of outcomes. The bubble caller assumes tech is still funded on the VC binary. That assumption is now wrong, which is why the bubble caller cannot hold both my views at the same time and I can.

The reformation breaks all three beliefs at once because it introduces something the old mindset does not know how to price. Finance and tech innovating together in ways that were not possible before. PE, of all places, is one of the clearest early signals of the shift.

PE Just Realized Every Model Company Is a Warrant Waiting to Be Written.

The last piece of the reformation is PE, which has been quietly written off by the market and the press. It has been a tough few years for PE with limited LP distributions and funds coming up to their maturity. A growing number of PE partners in recent months have been reaching out to me, and something is shifting. They are starting to see the same thing Microsoft saw when it looked at its enterprise distribution and OpenAI at the same time.

Microsoft did not just invest in OpenAI. Microsoft used its enterprise distribution to put OpenAI in front of the entire Fortune 500. Copilot inside every Office install. Azure inside every enterprise cloud budget. GitHub inside every engineering team. That distribution is what turned a research lab into the largest single-source revenue engine in enterprise AI history. Microsoft got the equity value of the scale that distribution created. OpenAI got the customer base it could never have built on its own. Neither side paid the other in cash. They paid each other in strategic access. That is a very different transaction than the venture capital "here is a check" model, and it is the transaction the PE partners I talk to are starting to see they can run.

Every large PE firm owns 50 to 200 portfolio companies (some even larger) covering hundreds of billions of dollars in end-customer revenue. That portfolio is a distribution channel larger than most public companies operate. Every model company needs distribution to grow into the revenue that supports the compute bills it is signing. Every PE portfolio company needs AI-first capability to compete with the AI-native disruptors already eating into their margins. The natural transaction is a warrant swap. The PE firm gives a frontier model company preferred access to its portfolio, priced right, with the operating playbook to make it stick. In exchange, the PE firm takes warrants in the model company. The portfolio companies get AI-first scale and the insights that come from running on frontier models under favorable terms. The fund gets venture-like returns from equity in one of the fastest-growing companies in enterprise software history without cutting a $500 million check upfront. That is a structure venture capital was never sized to write and cannot replicate. It is also very interesting to someone like me.

I have written the venture-side version of this in The Trust That Ate Venture Capital and the capital-cost version in The Most Expensive Money in the Room. PE, once written off as tired capital, has the exact scale, duration, and structure to do what VC cannot.

I could keep naming these structures for another 3,000 words. Debt facilities against compute pre-purchases. Long-duration equity vehicles built around ecosystem plays. Sovereign wealth co-investment structures. Balance-sheet finance for hyperscaler capex. Cross-portfolio data trusts. I picked a handful above to make the point. If you want to hear more, or if you know of a structure worth adding to the list, reach out. I have started looking at a few worth building inside Collective[i], and I am always open to a conversation with people who see this from a different angle.

Tech may be graduating Venture Capital.

Add the threads together. Big Tech is spending $725 billion in 2026 on infrastructure VC was never sized to fund. Compute is becoming a traded commodity with a futures curve. SaaS incumbents are trading lock-in for relevancy at exactly the wrong moment. PE is waking up to warrant swaps. Together these describe a financial architecture that has nothing to do with venture capital and everything to do with the arrival of real finance. Every prior tech bubble was funded by VC alone and ended when VC ran out. This one has the balance sheets of the trillion-dollar tech incumbents, the largest banks, the biggest PE firms, the exchanges, and sovereign wealth funds behind it. Those balance sheets do not run out on the same cycle VC does. This is a good thing. Sophisticated capital dampens the downside, extends the timeline, and lets the companies being built scale into markets nobody could have funded before. The bubble caller reads that shift as speculation. Anyone who has watched real finance work knows the opposite. Structured capital contains volatility. It does not amplify it.

But this article is not really about the bubble question. In fact this entire frame itself is the smaller story. It is one example of a much larger risk, and it is the one this piece was designed to make visible. Belief, magnified by AI, is what leads great investors, great leaders, and great boards to make mistakes at speed and scale that is shocking. Every doubter I talk to has a belief they built during a different era, a belief that was rewarded for decades, and a machine on the other side of every conversation that will validate that belief instantly and with a citation even when its completely wrong. Confidence at machine speed in a decision that no longer matches the world. That is the risk in every senior seat. And while I am always happy to take the bubble bet from anyone who wants to place it. This piece is aimed elsewhere. The leaders, operators, and boards who are already sensing what is happening need vocabulary to act on it. Sitting out because it feels like 1999 is going to look, in five years, like sitting out the Nasdaq in 2010 because it felt like 2001. Same emotion. Same cost. Same wrong answer.

The Skills to Manage the New Risk.

This piece was the setup. Belief/Gut magnified by AI is the deeper story, and the next three pieces cover the skills required to operate through it. Leadership. CEO. Board. Investor. None of those roles will look the same in five years because none of the assumptions they were built on hold. Each of the next three pieces covers one thing. What to change. What to keep. How to spot the belief traps before they land in an earnings report. If this piece gave you the frame, the next three give you the tools. Subscribe below and each lands in your inbox as it publishes.

What Comes Next

Post it on X, tag @smesser, and tell me where I am wrong. Share it on LinkedIn if the boards, investment committees, or executive teams you sit on need to see it, tag me at linkedin.com/in/stephenmesser, and tag the people who should be in the argument. Jensen Huang at Nvidia. Andy Jassy at Amazon. Satya Nadella at Microsoft. Sundar Pichai at Alphabet. Mark Zuckerberg at Meta. Dario Amodei at Anthropic. Sam Altman at OpenAI. Marc Benioff at Salesforce. Elon Musk at SpaceX. Larry Fink at BlackRock. Jamie Dimon at JPMorgan. David Solomon at Goldman Sachs. Steve Schwarzman at Blackstone. Marc Rowan at Apollo. Ken Griffin at Citadel. Terry Duffy at CME. This is the conversation every board, every investment committee, and every asset allocator needs to be having in public before the next capital allocation decision gets made.

If you are the person who just got this forwarded, welcome. Subscribe at reloadnyc.com. The archive covers the capital shift, the compute substrate, the AI-first operating model, and the motivated reasoning that keeps smart rooms wrong.

Connect with me on Intelligence.com for access to my entire verified network. If you and I both use it, we can share the parts of each other's networks that matter and stop the LinkedIn-connection theater every board and investor still runs.

Forward it to one person who still thinks this is a bubble.

Source · Studies, companies, and reporting cited above

The Reason Prior Tech Bubbles Broke Just Got Fixed.

Nobody Told the Bubble Callers. So What Does That Tell You About Belief?

Read more

Your gut was your edge. AI just turned it into a weapon of mass destruction.

The Agent Swarm Just Killed Scale as a Moat.

The West Just Got Destroyed. Your Company Is Next.

$65 Billion in Revenue. Seven Times Bigger Than a Year Ago. The Bubble Debate Is Over.