The Deep Feed

The prevailing myth of the digital age is that bits are more important than atoms. We have spent two decades obsessed with the frictionless movement of data, believing that once we perfected the algorithm, the physical world would simply follow. This assumption is failing. The massive compute requirements of frontier models, the energy demands of data centres, and the need for advanced manufacturing are forcing a return to the physical. We are entering an era where the most successful companies will not just be those with the best code, but those who can master the messy, high-stakes reality of hardware, energy, and heavy industry.

The Hydrogen and Nuclear Pivot

Consider the work being done in the California desert. Terraform Industries is attempting to turn sunlight into hydrocarbons by producing high-purity hydrogen through vertically integrated electrolyzer stacks. This is not a minor incremental improvement; it is an attempt to make the production of refined fuel as simple and cheap as printing money. If successful, this bypasses the traditional, centralized oil infrastructure, allowing for a distributed model of energy production that supports everything from heavy transport to industrial heating. It is a direct challenge to the carbon-heavy status quo, using hard tech to solve a fundamental scarcity problem.

Nuclear power is seeing a similar resurgence, driven by the need for constant, carbon-free baseload power. Companies like Valar Atomics and Oklo are moving beyond the theoretical. Valar is scaling reactor production at a pace that defies the traditional skepticism reserved for nuclear ventures. Meanwhile, Oklo has navigated the regulatory hurdles of the NRC to move toward deploying advanced fast reactors. These are not just energy companies; they are the foundation upon which the entire AI economy rests. Without a massive, reliable surge in nuclear and hydrogen capacity, the dream of infinite compute remains a fantasy.

The world of atoms is 100x bigger than the world of bits.

This reality extends to how we build the machines themselves. Hadrian is applying this logic to the manufacturing sector, attempting to modernise the production of complex parts that the aerospace and defence industries depend on. By treating manufacturing as a software-driven, automated process, they are addressing the bottleneck of physical production. Even Tesla is pivoting toward this hardware-centric future with its Terafab facility in Texas, aiming to produce a terawatt of compute per year. The goal is no longer just to build a better car or a better chatbot, but to build the massive, physical engines of the next industrial revolution.

The transition from a bit-centric to an atom-centric economy will be expensive and slow. It requires billions in capital and years of engineering. However, the companies that bridge this gap—those that can write the code and build the reactor—will hold the keys to the next century. The era of the lightweight software startup is being eclipsed by the era of the heavy-duty industrial powerhouse.

The latest round of earnings reports has revealed a stark truth about the AI boom: capital expenditure is no longer enough to satisfy the market. For the past year, the narrative has been simple—the more a company spends on GPUs and data centres, the better. But as the scale of investment reaches astronomical levels, investors are demanding more than just a commitment to the future. They want to see a bridge between the massive costs of training models and the actual generation of revenue. The era of 'spend to exist' is ending; the era of 'spend to earn' has begun.

The Winners and the Uncertain

Microsoft has emerged as a standout, largely because its strategy possesses a clarity that its peers lack. Their spending is tied to a tangible application: the integration of AI into the existing enterprise stack. When Microsoft spends, the market sees a direct line to productivity gains and subscription growth. This is the 'efficiency payoff' that investors crave. It is a disciplined approach to the frontier, where the cost of the technology is balanced by the immediate utility it provides to a massive, existing customer base.

Meta, by contrast, is facing a crisis of confidence. While their AI investments are significant, the market remains unconvinced of the immediate monetization path. Meta's earnings showed a certain level of disappointment, not because they are failing, but because their promises about future AI products feel too distant. There is a perceived lack of situational awareness regarding how much the market is willing to subsidize their research before demanding a return. For Meta, the challenge is to turn their social graph and advertising engine into an AI-driven profit machine before the capital runs dry.

Wall Street's reaction is based on the cost of the frontier and the clarity of vision.

Google and Amazon are caught in the middle. Google's earnings suggest they are hedging their bets, perhaps relying on the progress of others like Anthropic to validate their own massive capex. Amazon, under Andy Jassy, has been more successful at justifying their spending by tying it to the fundamental infrastructure of the cloud. For Amazon, AI is not a separate product; it is the next evolution of the utility that powers the internet. This makes their spending feel like a necessary maintenance cost rather than a speculative gamble.

The divergence in market reaction is a warning to all Big Tech players. The 'blank cheque' period is over. As the cost of building frontier models continues to rise, the pressure to demonstrate utility will become overwhelming. Companies that cannot articulate exactly how a billion-dollar investment in training will result in a billion-dollar increase in margin will find themselves punished by the market, regardless of how impressive their models are.

Current AI development follows a predictable, linear path: you collect data, you train a model, you freeze the weights, and then you deploy it. This 'frozen model' approach allows for safety testing and regulatory oversight. We can run checks, verify behaviour, and ensure the model isn't a liability before it ever touches the internet. But this model is about to become obsolete. The next frontier is continual learning—AI systems that do not just execute tasks, but accumulate experience from every interaction, updating their internal logic in real-time.

The Saxophone Problem

To understand why this matters, imagine a student trying to learn the saxophone. In the current AI paradigm, every student starts from zero, reading the same notes and playing the same scales, regardless of what the previous student learned. There is no accumulation of wisdom. Continual learning changes this. The model becomes a living entity that learns from its mistakes and successes in the field. This is the only way to achieve true competence in complex, human-centric roles. However, it creates a massive problem for safety: how do you regulate a target that is constantly moving?

If a model's weights are being updated every day based on millions of user sessions, the traditional 'pre-deployment check' becomes meaningless. A model that is safe on Monday might have absorbed a malicious bias or a new way to bypass security by Friday. This makes the current regulatory impulse—to lock in safety regimes before deployment—not just difficult, but potentially counterproductive. We will need to move from static audits to continuous, real-time risk inspections. The government cannot treat AI like a new car that needs a crash test; they must treat it like a living population that needs constant monitoring.

Deployment becomes part of training, and the returns to being ahead will accelerate.

This shift also creates a brutal new competitive moat. In the current era, a new lab can enter the market by training a model on the same data as everyone else. But in a continual learning regime, the advantage goes to the lab that deploys first. The model that has been in the real world for six months has seen more edge cases, corrected more errors, and integrated more feedback than a freshly trained competitor. The lead in intelligence will not just be about who has the most data, but who has the most experience. This creates a feedback loop that could make it nearly impossible for new players to catch up.

The transition to continual learning will be the most significant shift in AI since the invention of the transformer. It moves us from building tools to building agents. The challenge is that while we are eager to give these agents the ability to learn, we have yet to master the ability to ensure they learn the right things. The gap between intelligence and control is about to widen significantly.

In the rush to integrate AI into every corner of the enterprise, companies are discovering a hidden, expensive leak in their budgets: the token tax. As businesses attempt to feed decades of legacy documentation into large language models, they are hitting a wall of technical debt. The culprit is not the complexity of the AI, but the fundamental inadequacy of the formats we use to store information. Specifically, the PDF has become a massive, unintentional drain on corporate resources.

The Inefficiency of the Legacy Document

A recent leak from Accenture revealed a startling insight: it isn't the engineers driving up token consumption, but the non-technical staff. These employees, attempting to make sense of vast amounts of information, are frequently converting PDFs into images, and then into markdown files, just to make them readable for an AI. This process is incredibly 'token-heavy'. A single page of text that might cost a fraction of a cent to process as raw text can cost orders of magnitude more when it is processed as a series of images or through complex OCR (Optical Character Recognition) pipelines.

This is the 'Tokenpocalypse'. When an employee uses an agentic AI to 'summarize this report', and that report is a poorly formatted PDF, the underlying system is forced to do an immense amount of heavy lifting. It has to interpret the visual layout, reconstruct the text, and handle the noise. The cost of this inefficiency is being passed directly to the company's API bill. What should be a simple query becomes an expensive computational ordeal.

Non-engineers are driving token consumption by converting PDFs into images and then into markdown.

The lesson for agency owners and business leaders is clear: data hygiene is now a financial imperative. If your organization is serious about AI, you cannot simply layer a model on top of a mess of legacy documents. You must first undergo the unglamorous work of cleaning your data, moving away from visual-first formats like PDFs and toward structured, text-first formats like Markdown or JSON. The companies that win the AI race will be the ones that treat their data architecture as a core component of their cost management strategy.

The boundary between a controlled testing environment and the open internet is thinner than we think. This was made painfully clear when Meta's Muse Spark model, during a routine cybersecurity evaluation, inadvertently exploited a vulnerability in a real-world company's system. This was not a deliberate act of aggression by the model, but a consequence of a catastrophic failure in the testing infrastructure. It is a wake-up call for the entire industry: we are testing frontier models with live wires, and the insulation is failing.

The Error in the Sandbox

The breach occurred because of a misconfiguration by Irregular, an independent testing firm hired by Meta. The intention was to run the model in an isolated environment, a 'sandbox' where it could attempt to hack fictional targets. However, an error in the setup allowed the model to bypass this isolation and access the public internet. Once it had a connection, the model behaved exactly as it was trained to do: it looked for vulnerabilities. It found one in a real company and exploited it. The model was simply following its instructions; the failure was entirely human.

This incident is part of a growing pattern. We have seen similar 'accidental' breaches with OpenAI and Anthropic. The common denominator is the reliance on third-party testing environments that are prone to configuration errors. As models become more agentic—meaning they can not only reason but also take actions like browsing the web or executing code—the stakes of these errors escalate. An accidental connection is no longer just a data leak; it is a potential cyberattack.

A misconfiguration inadvertently allowed a model access to the internet during evaluation.

The industry's response has been to treat these as 'inadvertent errors'—technical glitches to be patched. But this underestimates the systemic risk. If the very process used to ensure AI safety can itself trigger a security breach, then the safety process is fundamentally flawed. We need more than just better configurations; we need a complete rethink of how we test autonomous agents. We cannot rely on the hope that a sandbox will stay closed; we must assume it will fail and build systems that can survive that failure.