"We were not the absolute masters of our universe; we lived in a world that was complex, confusing, and largely uncontrollable." - M. Mitchell Waldrop, The Dream Machine
In the early 1950s, a revolution was coming to life.
Engineers, physicists and psychologists were building the first computers in university labs. The technology evolved so fast that within a decade, computers went from room-sized machines built on vacuum tubes that could do about 5000 additions a second, to transistorized machines doing a million operations a second. The industry and technology were evolving so fast that most people inside of it were just trying to keep up.
When I was reading The Dream Machine, I couldn't help but notice the similarities to the situation that we're in today with AI. Back then, transistors suddenly made compute cheap and capable, doubling in power every couple of years. Today, intelligence, once expensive and rare, is now cheap and ubiquitous. Back then, new machines, architectures and languages showed up faster than anyone could master them. Today, it's new models, harnesses, tools and frameworks.
As they say, "history doesn't repeat itself, but it often rhymes."
Over the last few days, I've been spending more time thinking about the current state of our industry and what the future looks like when intelligence is abundant.
Are we in a bubble? Is SaaS dead? What will happen to developers?
Here's what I think.
Different kind of expensive
There is a lot of talk that the tech industry is a bubble. I disagree. I think the tech industry is healthy and growing well. In fact, I think this is just the beginning.
To start, we're already seeing the real-world impact that AI is having on our economy. The Fed reports that investment in information-processing equipment, software, and R&D contributed an average of 0.90 percentage points to annualized real GDP growth during the first three quarters of 2025, compared with 0.81 percentage points in 2000 during the dot-com bubble.
To be fair, these are pretty broad categories that are associated with AI, but the comparison shows that the level of investment in information technology is already economically significant and comparable to that of the dot-com era.
It's also hard to ignore the impact that the tech industry has had on the stock market. A handful of large technology companies account for a disproportionate share of US stock market gains. Just 7 stocks, known as the Magnificent 7, drove 53% of the S&P 500's 2024 return. Even more astonishing is that Nvidia alone accounted for 25% of the entire S&P 500 gain in 2024.
Usually, that kind of concentration in a single sector is seen as a risk. It was in the dot-com era, and that bubble crashed violently. So why wouldn't this one?
In my mind, there are a few key differences. First, at the height of the dot-com bubble in March 2000, 6 of the 20 largest domestic Nasdaq companies were losing money. In fact, 79% of the companies going public in 1999-2000 had negative earnings before their IPO. Many of these companies had no line of sight to profitability and were still dependent on outside funding. So when investor enthusiasm faded, these companies had nowhere left to turn.
Now compare that to the current situation where we have two cohorts of companies that are deeply investing in AI: established companies with mature cashflows such as Amazon, Meta, Microsoft, Nvidia and Alphabet, and upstarts (though I'm not even sure if you can call them that) such as OpenAI and Anthropic with high growth rates but more volatile cashflow. The established companies are already enormous, profitable businesses with obvious proven customer demand. Nvidia generated nearly $75B of operating cash flow in the first half of 2026 and Microsoft has nearly $76 billion in cash and short term investments. Meta ended 2025 with $81.6 billion in cash and marketable securities against $58.7 billion of long-term debt.
The upstarts are also impressive. OpenAI has roughly doubled its annual run rate in the last 10 months to an estimated $40 billion and Anthropic expects a second consecutive quarter of positive adjusted operating income (despite some interesting accounting). These companies aren't highly levered, like the ones in the 2008 GFC, nor companies that don't have proven customer demand. This ain't Pets.com.
Are valuations very high? Yes. All of the way down to the very young companies, but again, they aren't completely unfounded. AI companies are growing faster than ever - at least 15 gen AI startups have already crossed $100 million in ARR, some in under a year. Mercor went from $1 million to $100 million in annualized revenue in 11 months. Cognition went from $1 million to over $150 million in ARR in under a year, and has since climbed past $1 billion in run-rate revenue.
Then there's AI CapEx. Microsoft expects to spend roughly $175 billion on CapEx in 2026. Alphabet expects $195 billion and Meta expects $130 billion. It's not clear what OpenAI and Anthropic are going to spend but we know that both companies have secured enormous amounts of compute. OpenAI announced a deal with Nvidia for 5 gigawatts of capacity and Anthropic agreed to pay SpaceX $1.25 billion a month for compute. That's a lot of spending but the established tech companies have enough operating cash flow to finance it and the upstarts are growing so fast that it's not crazy to think that they can afford it.
It's also reasonable to read this and think, "well sure, they can afford it, but will it actually generate a return?" In the 2000s, Cisco was deploying a massive amount of infrastructure that was needed when Telecom usage exploded, but investors still overestimated how much infrastructure could be profitably deployed. I think this is a fair concern but the underlying dynamics look different this time around. By 2001, 95% of the fiber that was laid sat dark. Industry had outpaced demand and deployed infrastructure that wasn't utilized. Today's labs are running into the opposite problem - they are routinely compute-constrained. A skeptic would point out that while AI companies are compute-constrained today, eventually the compute supply will catch up and at that point the AI companies need to be cash-flowing machines otherwise we're in for a bad time. But I think this misses the other vector that's growing in magnitude and will contribute to compute demand - companies that are moving off of the frontier labs and running their workloads on open source models. This vector is still early but it's growing quickly and I expect over the next 3-5 years will start to account for a good chunk of compute demand. So until we start seeing signs of compute saturation, I'm inclined to believe that there is still headroom to grow.
And even if sentiment turns, there's the US government backstop. The tech industry, and specifically AI, has never been more important to our national security. In October 2024, Biden directed the government to strengthen America's AI capabilities. Trump reinforced this plan in 2025 and framed AI as a race with economic and military consequences. The CHIPS Act shoveled $52.7 billion into the domestic semiconductor industry including nearly $40 billion in manufacturing incentives. The US government even took a 9.9% stake in Intel. We've also seen the private market push this narrative further, with companies such as Palantir taking a strong stance on Sovereign AI.
At the same time, open-source models from China have flooded the market as US-based companies are looking for ways to save on their ever-increasing AI spend. Despite OpenAI and Anthropic releasing new models that continue to push the frontier of cost/token, it seems like China is hell-bent on trying to undermine US-based AI companies by offering free models that perform at near frontier-levels of intelligence.
In this sense, the US-based AI companies are essentially fighting a proxy intelligence war with China. And if we know anything about the US government and proxy wars - it's that they don't mind funding it for years if it serves the national interest.
Allowing this industry to collapse would undermine the previous and current administration's national strategic initiatives and risk giving China the lead in a technology that is already defining economic and military power for decades. I think it's more likely that these companies start to get effectively nationalized than the US government allows them to fail.
So where does all of this leave us? Clearly we're in a period of massive change in our industry but I'd wager it's unlike any other "bubble" we've seen which is why I hesitate to call it a bubble. Are there some perverse incentives driving investors to keep jacking up the paper value of some AI companies? Absolutely. But there is also clear customer demand, growth rates unlike anything we've ever seen before, and minimal leverage. So will all of this seemingly crash one night? I really don't think so. At least not anytime soon.
When intelligence gets cheap
Every decade or so, something scarce becomes cheap and it drives the next generation of startups. In the mid-2000s, AWS launched compute primitives in the cloud and all of a sudden, building a company got cheap. No physical servers needed, just an AWS account and you're off to the races.
The result was an explosion of software companies that took a business workflow, encoded it in software, and tried their best to sell it better than anyone else. From marketing to sales to operations to IT to finance, every function was SaaS'ified.
Most of these companies did more development than research. They spent countless engineering hours building integrations, dashboards and reports. Most ML teams were focused on which item to recommend, which emails to send, and which discount would push someone to check out. Teams of PhDs spent years optimizing customer CAC and LTV (which isn't necessarily a bad thing) while the product itself barely changed (definitely a bad thing).
But general intelligence changes things. I don't need a SaaS product to write, schedule and post my social media posts or project manage my work anymore, I can prompt Claude or Codex and it just works. Between generalized intelligence models from OpenAI and Anthropic and teams building more of their own software, the app layer is becoming increasingly hostile to new entrants.
As it does, history is rhyming and the pattern is clear. Cheap cloud compute gave us the SaaS era, and cheap intelligence will define the next one. Any startup can get frontier-level intelligence at a fraction of the cost and it's only getting cheaper. So if everyone has access to the same intelligence, how do you differentiate?
I think the answer is investing in research as much as development.
Some companies are already doing this. Ramp launched Ramp Labs, an AI research group experimenting with everything from model routers to RL to new spreadsheet products. Cursor has been training its own models from the beginning with the Tab Model (shout out to Supermaven) to the current Composer Models to improve their coding workflows. Stripe built a Payments foundation model that increased their authorization rates by 3.8%, which makes a big difference when you're processing trillions of dollars. Even more established enterprises such as Thomson Reuters are building their own LLMs in-house on open-source models.
These companies are combining their proprietary data sets with custom models, harnesses and tools to build products no one else can.
The key distinction here is that we have to start treating generally intelligent models like ChatGPT and Claude as tools that help accelerate your internal R&D to build better, smarter products, not as something that you slap onto your dashboard.
I don't need a health wearable agent I can talk to, I need a health wearable that is able to proactively track more discrete signals, at a granular level, and tell me that I'm at high risk of some condition due to the hardware and software advancements the R&D team has made by leveraging AI.
And this applies to companies in pretty much every industry. We should have smarter, faster, and better defense systems, manufacturing systems, energy systems, construction processes and everything else because teams pushed AI to its limits to develop better products.
Investing in research isn't just the right product strategy but also the best recruiting strategy. The reality is that most startups can't compete with OpenAI and Anthropic on cash comp - top AI labs pay roughly three times what most companies can.
But you can compete with them on mission and values. The best researchers and engineers want to deeply believe in a mission. Anthropic is living proof of that. $2 trillion of market cap will soon exist because some researchers didn't believe that OpenAI was living its values.
But you have to give them something to believe in. And optimizing another checkout flow ain't it.
So if you're starting a company today, I think the playbook is to build something where the research is the product. Especially in the physical world, there are so many opportunities to improve the life of millions of people through fundamental research. Better batteries, medical devices, solar, wearables, consumer hardware and so on. And it's never been cheaper and easier to experiment.
That's where I think the next trillion-dollar companies will come from.
The mongrel and the model
For some developers, AI feels like an existential crisis. Claude and Codex are now good enough to write most of your code, if not all of it. If you spent the last decade, or longer, getting really good at writing code, watching a model do it in seconds is a strange experience.
For others, it's a tool straight from science fiction. One that can release them from the drudgery of hand typing code like a mongrel and unlock their deepest technical desires and aspirations.
Either way you look at it, the role, toolset and expectations are changing.
With the shift towards more research-based work, I think most companies will, and should, expect that developers can stand up an open-source model, fine-tune it on proprietary data and evaluate it, and serve it cheaply. I think this is going to be as core to the most valuable developers as shipping a web app was in 2015.
I routinely fine-tune OSS models using Codex and eval them against our existing pipeline. Just as I would test a code change, I'm testing a model change. And to be honest, it's a lot more fun than building yet another queue.
What I'm confident about is that good developers aren't going anywhere. The world doesn't need fewer people who can solve hard technical problems. If anything, with research becoming the differentiator, it needs more of them. The job is just changing shape.
The developers who treat the hours they're no longer spending on code as time to learn how the whole stack works, from the model up, are going to be the ones building the most interesting things of the next decade.
The quote at the beginning of this blog by M. Mitchell Waldrop has been bouncing around my head for a few weeks now: "...complex, confusing, and largely uncontrollable." That was true of the people building the first computers, and it's true of us now.
Imagine living through it. Punch cards to transistors in the 1950s. Time-sharing and integrated circuits in the 1960s. The microprocessor and the first personal computers in the 1970s. Each decade making the last one look primitive.
The people living through that time never had the luxury of feeling settled. And neither do we. They just kept building anyway, and it turned into the internet.