The AI "Bubble" Needs to Accelerate

Every time a technology chart goes vertical, the bubble chorus starts singing. It happened with the railroads in the 1880s. It happened with the internet in the late 1990s. And now it is happening with Generative AI.

The narrative is seductive: too much money chasing too few ideas, a collapse is inevitable.

But the underlying data tells a far more complex story. When you look at who is building, how it is being funded, and what the capacity is actually being used for, the AI buildout is not necessarily a bubble.

It is true that there may be some over-exuberance, some misallocated resources, and misplaced bets, as there is in every growth market. However, there continues to be a supply-side crunch driven by the ever changing needs of AI (hint - changing cooling needs and new generations coming out every 6-months).

This is not another tech boom. It is the early industrialization of cognition: a capital intensive, energy constrained expansion that lowers the marginal cost of reasoning itself. Wall street analysts have no clue how to price in how true machine intelligence can reshape our economy quite yet.

And unlike the internet era, this buildout requires many scarce inputs at the same time. We need chips and we need energy. We also need more research and to use machine intelligence to create even more intelligence. We have crossed the Rubicon in the quest for AGI - that march can't be turned around.

There are likely to be lots of people who made the wrong bet on the wrong technology at the wrong time - and that capital will be wiped out. Yet the fundamental principle behind “scale-pilling” remains the same - more compute means smarter models which (eventually!) turns into GDP when the dust settles.

Let's dive into the data.

A Brief History On Scale Pilling

By history, I mean 2023/2024, which we call “The Age of Scaling”. Popularized by Dylan Patel on the Dwarkesh podcast and lots of researchers like Ilya Sutskever and Rich Sutton's 2019 “The Bitter Lesson” argued that compute trumps clever algorithms. A March 2023 X thread by Andrej Karpathy went viral, quipping “scaling is the new black” after GPT-4 demos. This fed into broader discourse - the fundamental belief was that we cannot afford not to build build build.

And no one wants to get left behind in this equation.

This is where the industry's psychology began to shift. For over a year, “just scale it” was the dominant mindset, a kind of inevitability narrative. The moment serious voices questioned that trajectory, it created a wave of uncertainty: if scaling isn't the only lever, then the economics of the entire buildout need to be re-evaluated. That's the backdrop for why markets reacted so violently.

Which is why Ilya's comment on Dwarkesh last week hit so hard: “Scaling is not enough”. That and Nvidia falling despite beating earnings and a very bearish market with Michael Burry headlines calling a bubble has led to today's climate.

What everyone is missing is that two things can be true:

1. Machine intelligence does not scale linearly with cluster size (aka, there are indeed, diminishing returns).
2. AI will embed itself in every corner of the economic machine and that means more and more compute will be needed (even if the compute becomes more efficient).

Railroads, Fiber, and GPUs

To evaluate bubbles, ignore the hype. Look at CapEx as a percent of GDP, the real supply constraints, and the long term economic impact.

• We have already surpassed Dot com era infrastructure intensity.
• We are now moving toward railroad era intensity, but with far tighter physical constraints and far more immediate utility.
• And unlike fiber with DWDM, there is no way to magically multiply GPU capacity.

Infrastructure Intensity & Outcomes Across Eras

* Economists use a 3–5 year window because that's the typical productive life of technology capital and the historical timeframe over which new infrastructure translates into measurable GDP gains. GDP Figures From Investment Research Partners.

The takeaway: we have already surpassed the Dot-com infrastructure intensity and are approaching “railroad” territory, but with much tighter physical constraints and far more immediate utility.

Demand: Forecast vs. Backlog

The single biggest difference between 2000 and 2025 is how demand shows up.

2000s: Forecast Driven Fiber

In 1999, Global Crossing and friends trenched fiber across oceans based on models that said internet traffic would double every 100 days. That story collapsed when DWDM let operators pump 100x more data down the same glass. Supply became effectively infinite. Bandwidth prices fell through the floor. Utilization on many networks sat near ~1–3%. They built the highway before the cars existed, and then discovered the highway itself was infinitely expandable.

Current Day: Backlog Driven Compute

Today, the sequence is reversed:

• Silicon is pre-sold. Nvidia's Blackwell generation is effectively sold out for 12 months. That's not “pipeline interest”; those are firm orders from hyperscalers who are fighting each other for allocation.
• Capacity is pre-leased. In core US markets like Northern Virginia, vacancy rates are under 1%, and ~70%+ of capacity under construction is already pre-leased before the slab is poured. Many colo projects do not even break ground until they presign usage with credible consumers and also sign the energy leases.

Why You Can't Multiply a GPU

The telecom bubble popped because of a single technology (Dense Wavelength Division Multiplexing, DWDM) which let operators multiply network capacity without deploying more physical capital. The bottleneck disappeared overnight. Prices followed.

There is no equivalent “free multiplier” for AI compute:

• You cannot push a firmware update and make an H100 do 100x more FLOPs.
• Scaling requires new wafers, new CoWoS capacity, new HBM, new racks, and new megawatts. All of that costs real money and takes real time. There are emerging architectures—Cerebras wafers, Recogni accelerators, TPUs, and other domain-specific chips—that may meaningfully reduce compute-per-watt. But those require new deployments, not upgrades to existing silicon. None have yet demonstrated the kind of 10–100× supply-side expansion that DWDM brought to fiber.
• Lead times on high-voltage transformers are 3+ years. Fabs take most of a decade. Transformers, switchgear, and cooling are all very real bottlenecks. For the operators we work with at Aravolta, the new currency isn't just megawatts, it's speed to market. In today's market, the first operator to stand up a liquid-cooled, power-dense hall doesn't win a contract, they win a waiting line. That's exactly why operators are turning to platforms like ours to accelerate provisioning, automate site readiness, and reduce the time from “hardware delivered” to “revenue-generating GPUs online.”

There is a very real scarcity floor: In 2000, capacity could explode with a software and optics upgrade. In 2025, capacity expands on 24–36 month cycles tied to fabs, power, and grid upgrades.

TPUs are not the only Innovation

The big question underneath all the NVIDIA discourse is this: If scaling is slowing and demand is exploding, where does the next supply-side breakthrough come from?

While the industry loves to fixate on Google's TPUs (especially with Gemini being trained on them), Google has more incentive to make Gemini leaps and bounds better than instead making TPUs a broadly available commercial product. That leaves a wide strategic opening.

And this is where things get interesting: behind the scenes, a wave of alternative architectures is emerging—wafer-scale engines, LPUs, reconfigurable dataflow chips, RISC-V chiplets. These companies aren't just “competitors”; they represent the only plausible path to a 10×–30× improvement in compute-per-watt. If one of them hits, the entire supply chain shifts. That's why it's worth looking at who's real and what they're actually delivering.

No one can maintain a lead forever - though it remains to be seen who the big winners are and what the true time scale is.

What this means for the buildout also remains to be seen - will we still need liquid cooling? Move to immersion? Will the energy needs decrease significantly with chip innovations? Or will we just deploy more and more scale?

When Would This Become a Bubble?

To be fair, this can still go wrong. But “AI mentions on earnings calls” or “Nvidia's stock multiple” are not the metrics that matter. A real AI infrastructure bubble requires at least two of these structural breaks.

Furthermore, colocation markets provide a strong signal on enterprise activity.

Key Primary Markets: Inventory and Vacancy (H2 2024)

Notes: Northern Virginia remains the largest market. Atlanta led absorption in 2024. Vacancy rates are record lows across primaries (overall 1.9%). Sources: CBRE H2 2024 Report.

Across every market signal—vacancy, absorption, rents, and power availability—the story is consistent: we are structurally short on capacity, not long. There is no sign of oversupply anywhere in the stack.

The Bubble Dashboard

None of these conditions are here today. All of them point in the opposite direction. If you look at the Q3 2025 earnings prints, the “Revenue Expansion” phase has officially begun. We aren't just buying chips; we are selling intelligence.

• Cursor: Fastest SaaS company ever to reach $1B ARR (24 months)
• JPMorgan: AI coding tools boosted productivity by 20% across 60,000+ technologists.
• Walmart: AI saved 4 million developer hours, equivalent to ~2,000 full-time work years. Route optimization eliminated 30M unnecessary miles and 94M lbs of CO₂ using AI routing.
• Meta: AI-powered “Advantage+” (Ad Engine) now driving a $60B+ revenue run rate, achieved both +14% ad impressions AND +10% higher ad prices
• Adobe: AI-related revenue exceeded $5B; Firefly AI-First ARR approaching $500M.
• Salesforce: Closed 5,000 Agentforce deals in just months as enterprises upgrade from old chatbots. Internal data shows the system achieving an 84% resolution rate with only 2% human escalation.
• Duolingo: Launched 148 new AI-generated language courses in one year, which previously was a decades-long workload. “Duolingo Max” AI tier increased ARPU by 6%, 51% user growth.
• ZipRecruiter: AI cut time-to-hire by up to 30% and boosted screening efficiency for 64% of recruiters.

The Two Market Reality: Froth vs. Bedrock

We need to stop treating AI as one monolithic asset class. There are two distinct markets here, and only one of them is showing bubble behavior.

Layer 1: The Consumer Layer

This is where skepticism makes sense. Thousands of startups raised money in 2023 to build simple interfaces around foundation models. Most of them will not survive.

• Many analysts expect 85 to 90 percent of these companies to fail within three years.
• Thin moats and high churn make this layer extremely competitive.
• These companies often depend entirely on someone else's models and infrastructure.
• The foundation model providers are moving up-stack, watching what works on their own ecosystem and then building it themselves: the Amazon Basics playbook.
• They see the usage data, they see which workflows take off, and they just… ship their own version. The playbook is simple: startups test the market, OpenAI/Anthropic/Google watches what sticks, and then ships the “Amazon Basics” version at platform scale. Rewind, Adept, Jasper and countless others learned this the hard way.

But this layer is not just noise. It is already producing durable winners.

• New tech giants like OpenAI and Anthropic are here to stay.
• A new generation of AI native software companies is showing real revenue and real value. Examples include Cursor, Lovable, and others that are scaling faster than early SaaS leaders.
• Just like the Dot com era, many companies will die, but a few will become generational giants.

Even the failures in this layer help the system. Every failed startup has already spent its venture dollars on GPU compute. They end up subsidizing the infrastructure layer.

Layer 2: The Infrastructure Utility Layer

This layer is completely different. It includes the data centers, the power, the cooling, and the silicon. It behaves like a utility, not a speculative bet.

• The spend becomes a recurring operational cost, similar to electricity or cloud storage. That stickiness is already visible in Aravolta's own data. Once an operator instruments their power, cooling, and asset lifecycle workflows with us, they don't rip it out. They expand it. AI infrastructure behaves like a utility stack, and the operational software that sits on top behaves like one too.
• Infrastructure providers benefit from long term, sticky usage patterns across industries.

National Security and the Race to Sovereign AI

If the private market slows, the government becomes the buyer of last resort for AI infrastructure. We've seen more examples of this popping up, from the Genesis Mission (likened to a modern Manhattan Project) to the massive federal mobilization of cloud resources.

It seems that talks of climate change have completely fallen off the public agenda, replaced by a rush to own the intelligence layer. Project Stargate and the recent AWS federal commitments position compute not as a commercial commodity, but as a national security asset.

This creates a hard demand floor that does not depend on consumer hype.

The Sovereign “Put Option”

While VC Twitter debates whether chatbots have product-market fit, the federal government is effectively underwriting the physical buildout.

• The $50 Billion Anchor: AWS recently announced a $50 billion investment specifically for US government/defense regions (Top Secret, Secret, and GovCloud). This involves 1.3 Gigawatts of capacity, roughly a nuclear reactor's worth of power, dedicated solely to national security.
• The Genesis Mission: An Executive Order has launched a “Manhattan Project” for AI, directing the DOE and National Labs to build “scientific foundation models.” This isn't a startup looking for revenue; this is the state nationalizing the science layer of AI.

However, simply having the government write checks for power doesn't guarantee efficiency. If the US is serious about accelerating AI capacity, we don't just need more chips and more megawatts; we need the software layer that turns physical data centers into productive AI factories.

We are moving from an era of “build at all costs” to “operate with precision.” Whether the buyer is a hyperscaler or the Department of Defense, the need for truth is the same. They need to know if that hardware is being utilized effectively, or if it's burning out due to thermal stress.

That's the layer Aravolta is building. We provide the telemetry that turns compute into a measurable, performant investment.

Conclusion

When you zoom out, this does not look like the Dot com era. It looks like the start of a new industrial revolution. It is slow, physical, expensive, and inevitable.

The question is not “Is this a bubble” but “How do we get more compute, more megawatts, and more useful work out of every cluster before someone else does?” That requires a new kind of operational software—the layer that helps GPU data centers extract more useful work from every megawatt. That's exactly the layer companies like Aravolta are now building.