Today we will cover the recent Epoch AI report commissioned by Google DeepMind on the primary trends we are seeing in compute utilization for AI and the direction we are heading by 2030.

This is quite an important bet to make since being directionally correct over the next four years is likely going to be a “make it or break it” moment for many in the long term.

The key takeaway

For tech sales: While we often try to focus on more obvious indicators when it comes to identifying the right tech companies to work for (financial metrics, sales reps feedback and competitive performance), reasoning from first principles means identifying the overall trend in the industry. The biggest challenge over the next year is not whether AI will continue to scale but rather what are the unique breaking points where traditional strategies do not seem to work and there is a need for creative solutions. We've witnessed some of it today with coding tools, which have so far been the most lucrative growth area together with accelerated computing hardware. Being ahead of the game means looking for companies that are leveraging AI and the significant VC funding on the market to actually go after solving "too difficult" problems. If you pay attention first to the direction that researchers and scientists are taking, you can see where their peers are starting to build those new companies. Practically this means trying to identify companies that reduce complexity rather than manage it and can eliminate multiple problem categories rather than incremental improvements to existing models. The most difficult part remains having a good sense of market timing.

For investors: There are a number of reasons why we might be able to keep scaling on a similar pace as before, however it's important to note what this implies. If scaling is "commoditized", then it's quite possible that meaningful breakthroughs in value creation would be tied to either proprietary data moats (or more importantly the capability to generate such data) or brand new technologies that flip the energy constraints on their head. In the context of cloud infrastructure software, most probably that means that just having the most servers doesn't guarantee "winning". If AWS won by making self-hosted servers irrelevant, then the next AWS will likely make the current cloud hyperscalers' abstractions irrelevant. In practice that means that an interesting new company would not simply do Kubernetes (but a bit better), but would introduce technology that makes K8s unnecessary (while still allowing laggards to integrate). This is already happening in a variety of scientific contexts, the question is how quickly we see some of that talent move into the intersection of computing software and hardware. The wild card is regulatory risks, as already seen by the almost self-sabotaging choices made by EU leaders.

Scaling and capabilities

Compute scaling has played a key role in AI development, and will likely continue to do so.
Compute for training and inference drives improvements in AI capabilities, and much progress in AI research has come from developing general purpose methods to enable the use of more compute.
The trajectory of AI development can be forecasted based on continued compute scaling.
Scaling has significant implications across many areas of AI development: training and inference compute, investment, data, hardware, and energy. When we predict that compute scaling will continue, we can then examine the consequences within each of these — and how they need to scale accordingly to allow compute scaling trends to continue.
Exponential growth will likely continue to 2030 across all key trends.
Across training and inference compute, investment, data, hardware, and energy, we argue that a continuation of existing trends is feasible. We explore each factor in detail, showing how growth could continue to 2030, and discussing the most credible reasons for slowdown or acceleration before then. We argue the most credible reasons for a deviation from trend are changes in societal coordination of AI development (e.g. investor sentiment or tight regulation), supply bottlenecks for AI clusters (e.g. chips or energy), or paradigmatic shifts in AI production (e.g. substantial R&D automation).

The big question when making a bet over the next few years is not whether compute and data needs will grow (this is obvious and self-evident); it’s rather whether we continue to scale on a similar growth curve on the uptrend. It’s important to remember that already this year we’ve had two big FUD (fear, uncertainty, doubt) campaigns around this topic.

“Selling Enterprise AI” 2.8 is out, focusing on the recent speech by the Replit CEO Amjad Masad at YC on why SaaS services will become increasingly commoditized.

Subscribing today will give you access to more than 5 hours of Enterprise AI content, weekly calls, and a private version of The Database.

The first one was when DeepSeek was released (“super bearish for NVIDIA; anybody can train a super-smart model on peanuts”), and more recently, concerns related to scaling performance for GPT-5 (“it’s bad, we’ve peaked”). DeepSeek obviously pushed reasoning into the mainstream, which led to a massive explosion of usage for both OpenAI and Anthropic models (3–5x ARR within nine months), and GPT-5 had a rocky product launch that is now turning into an aggressive pivot on the agentic coding side (Codex is being perceived as a better performer than Claude Code).

Why compute rather than algorithms or data?
There are two common objections to a scaling-focused view of AI progress: algorithmic innovations and data. We argue that although they complicate the picture, they remain compatible with it.
Algorithmic innovations play a vital role, but they are closely paired with compute scaling. To paraphrase the Bitter Lesson, the most important and effective algorithmic innovations are general-purpose methods that enable compute scaling. Moreover, there is some evidence that algorithmic innovations rely on compute scaling for their development. This suggests that we should anticipate algorithmic progress, but enabled by, and focused on, compute scaling. Nevertheless, this is a key uncertainty. Capabilities could improve faster than predicted here, if compute is not a bottleneck.
Data is essential for AI training, and the quality of datasets can significantly influence results. However, there are two reasons to think that compute is more of a rate-limiting input. First, compute is more of a bottleneck in the current paradigm of AI training, at least for general-purpose LLMs. We could scale up for at least a few more years using existing public text data and other modalities. Second, it appears increasingly likely that inference scaling will make training more compute-intensive, effectively using compute to generate data for reasoning training (Data won’t run out by 2030, although human-generated text might). Specific data bottlenecks can be important within particular applications, and we discuss these further in Capabilities in scientific R&D. Hence, we must consider data availability when we investigate scaling, but this remains compatible with a scaling-focused view.

In traditional software, the most significant leaps in performance often occur because of code and data footprint optimization. Training new frontier models has nothing to do with traditional software creation, which is about explicitly programming rules and logic.

Training a new frontier model is about creating new neural networks, which is data-driven, and the "logic" emerges from training on examples rather than being coded. When someone tries to apply historical software trends to scaling neural networks, they are fundamentally making bad-faith arguments. The only truly reliable observation we have on scaling AI comes from "The Bitter Lesson" essay.

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. The ultimate reason for this is Moore's law, or rather its generalization of continued exponentially falling cost per unit of computation. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. These two need not run counter to each other, but in practice they tend to. Time spent on one is time not spent on the other. There are psychological commitments to investment in one approach or the other. And the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation. There were many examples of AI researchers' belated learning of this bitter lesson, and it is instructive to review some of the most prominent.
…
One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.

We have seen this play out repeatedly in the last few years as companies try to build capabilities that the models can’t handle, only to see them run perfectly with the next model release. In other scenarios, companies have also built what’s called “scaffolding”—basically tools and frameworks to force desired behavior—which was not working, only to see months of R&D become obsolete because the next frontier release was able to achieve the task “out of the box.”

What can’t compute scaling predict?
What this doesn't allow is predicting when we have general intelligence, i.e. AI that can perform any cognitive task at the level of a skilled human. This question suffers from two massive uncertainties: gaps in AI benchmarks and our understanding of them, and gaps in current AI capabilities that might not be filled in the next five years.
Current benchmarks might not adequately represent the most difficult-for AI tasks that humans do. And not all benchmarks show AI progress yet: there are tasks where AI doesn't yet show much improvement with scaling so far, for example "autonomously prove a new substantive mathematical theorem". It is fundamentally uncertain where AI will have reached expert level performance by the time it solves existing benchmarks (that show progress). It is also fundamentally uncertain when AI will solve all existing benchmarks, because it only shows progress on some of them.
Nevertheless, it is fairly certain that AI will solve many challenging benchmarks by 2030, and these have clear implications for useful tasks that AI will be able to perform.
Meanwhile, gaps in the capabilities of present-day general-purpose AI systems shed some light on the capabilities that AI could fail to achieve by 2030. AI models excel at identifying relevant information from a large training corpus, but frequently veer into illogical hallucinations. They are brilliant at ingesting large amounts of data and identifying underlying patterns, yet fail to reliably apply reasoning steps that would seem natural to a human. Reliability and robustness are problems more broadly, although they have at least shown incremental improvement with scaling. AI excels at solving closed-ended optimisation problems such as games, yet struggles to perform consequential actions in the real world with agency. It can perform shallow processing of long content much faster than a human, but it struggles to use this long-context information for solving challenging problems. There are enough gaps in current AI capabilities that it is hard to even be certain which of them are overlapping – perhaps long-context comprehension is related to robustness of reasoning, or perhaps they are entirely separate problems. These limitations are related to the challenges with designing and interpreting AI benchmarks: adequately benchmarking these limitations is also an open problem.
It is uncertain which of these AI limitations will improve by 2030, and by how much. It is uncertain whether these will improve by "just scaling" existing systems with small modifications, but also, it is unclear how much compute they would need. Consider the example of reasoning models: pre-existing systems already used inference scaling, but reinforcement learning (RL) made this far more effective, yielding breakthrough results in several benchmarks. Does this challenge the view that scaling is the driver of progress? Arguing in favour of scaling-driven progress, many researchers predicted ahead of time that better inference scaling would be necessary, and this arrived after models scaled up sufficiently for reasoning RL to work. Furthermore, training scaling holds for RL: using more RL training compute improves the capabilities that reasoning models achieve.
On the other hand, this emphasises the challenges of prediction from existing results. The areas where AI struggles today can sometimes see breakthrough algorithmic progress, and this is inherently hard to predict.

The reason why it’s critical to pay attention to what researches (i.e. AI scientists) are talking about is because fundamentally the biggest leaps will occur as scientific discoveries and will often be optimised for those types of problems. We’ve only experienced a similar type of setup with the space race in the 1960s, which led to a number of technologies being created and then productised over the next decades (portable computers, satellites, solar panels, medical imagining tools, air and water purification, cordless power tools, digital cameras).

An obvious example is the focus on mathematical reasoning. Creating models that surpass every single human in problem solving capability has multiple downstream effects that will be productized in corporate use cases. If we narrow down a couple of scientific categories:

Software engineering: Many of today’s day-to-day tasks are likely to become automatable by AI agents. Existing benchmarks based on well defined software issues, such as SWE-bench, are on track to be solved in 2026. Current progress on solving defined hours-long scientific coding and research engineering problems (RE-Bench) is slower, but on its current trajectory would be solved in 2027. A key uncertainty is whether human supervision will be a bottleneck for more open-ended problems.
Mathematics: Challenging mathematics reasoning benchmarks, such as FrontierMath, could be solved as early as 2027 on current trends. Mathematicians predict AI capable of solving such benchmarks might help them by developing sketch arguments, identifying relevant knowledge, and formalising proofs. This would allow AI to fulfil a similar role in mathematics to coding assistants in software engineering today. Even more than for software engineering, a key uncertainty is whether existing mathematics benchmarks are valid for predicting such capabilities. The most challenging mathematics benchmarks today are further from mathematicians’ day-to day work than software benchmarks are from that of software engineers. It is unclear when AI can rise to the level of autonomously proving substantive results, but it is plausible that this will happen before 2030.
Molecular biology: Public benchmarks for protein-ligand interaction, such as PoseBusters, are on track to be solved in the next few years, although the timeline is longer (and uncertain) for high-specificity prediction of arbitrary protein-protein interactions, especially further from training data. Meanwhile, AI desk research assistants are set to help in biology R&D in coming years. Open-ended biology question answering benchmarks are on course to be solved by 2030, albeit with large uncertainty. Importantly, advances in basic biology R&D are likely to take several years to lead to downstream changes in e.g. pharmaceutical development, due to bottlenecks in both wet lab experiments and clinical trials.
Weather prediction: AI weather prediction can already improve on traditional methods across timescales from hours to weeks. Moreover, AI methods are cost-effective to run, and are likely to improve further with additional data. The next big methodology challenges lie in improving prediction calibration at current horizons, rather than extending them further.
There are outstanding improvements to be made in two areas in particular: forecasting rare events, and integrating additional data sources. Using more historical data and more fine grained historical data for training can improve predictions, and more real-time sensor inputs could be integrated for better performance in deployment. There are important challenges in development and deployment: funding the research, getting access to data (particularly at low latencies in deployment), and in some cases even permissions to install data recording equipment. Nevertheless, improved weather prediction methods could achieve significant benefits in the wider world, helping in areas such as power infrastructure, agriculture, transport, emergency response, and everyday planning.

If we continue to progress on mathematical reasoning at a similar pace, we should already be seeing, within the next 12 months, an increase in scientific discoveries directly related to AI.

There are many different applications in biology where AI is already showing great promise. AI is already being investigated across many areas: prediction of biomolecular structures and interactions, analysis and editing of genomic data, imaging, lab robotics, and more. Due to the breadth of the field, we focus on two key areas that represent the divide between specialised tools and general-purpose agents: AI for biomolecule prediction and design (particularly proteins), and biology desk research. Other research areas could be vitally important, but are either less straightforwardly within the purview of AI (such as robotics and imaging for wet lab research), or less straightforward to analyse. AI for prediction and modelling of biomolecules has seen staggering success. AlphaFold2’s principal authors recently shared the 2024 Nobel Prize for Chemistry.
AI approaches such as AlphaFold have revolutionised protein structure prediction, achieving near-experimental accuracy for many well characterised protein domains in their equilibrium state. Subsequent work has attempted to carry these successes over to other problems, such as other biomolecules like DNA/RNA, dynamic structure, molecule interaction, and even target identification and protein design.

The combination of agentic abilities (helpfully advanced by progress in coding agents) and reasoning based on discovering new information (mathematics) is being applied across various scientific domains, with biology/chemistry experimentation for medical purposes being the first frontier to see material progress.

Ultimately, the goal of AI researchers is to create an agentic version of themselves that can be deployed at scale to train new models. Similar approaches are likely to be adopted across various scientific environments.

More importantly, since these developments are occurring simultaneously, there are likely to be unexpected multipliers arising from parallel discoveries in different adjacent areas. If we pull back to tech sales, value capture for the majority of this activity will occur at the bottom of the stack (and as a general societal benefit).

The computing needs to train these new capabilities and later deploy them for inference at a massive scale are, well, exponential. My advice, as always, remains to either focus on positioning yourself in cloud infrastructure software in a way that benefits from this explosion in new workloads or refocus on the impact this will have in real life, i.e., real-world AI implemented in deep tech.

If you are looking at this and still thinking that your mid-market role at Workday is an “opportunity,” you are fundamentally not making the right directional bet.

The Tech Sales Newsletter #111: AI in 2030

The key takeaway

Scaling and capabilities

The Tech Sales Newsletter #110: The Larry Supremacy