The Commoditization of LLM's and Advent of AI Agents: How Hardware Constraints Drive Software Innovation
AI disruption has been front of mind for the investment world for much of the past two years. Since its widespread release and initial surge of attention, GenAI has quickly gone from an emerging technology to a cornerstone of business development. While much of the wider investment landscape has lamented the lack of a clear return on investment for major hardware budgets at cloud infrastructure companies, there have been differentiated participants who have taken strategic and specific approaches to finding business applications, and are very much realising lucrative returns on their products that drive material value.
In our past comments, we have outlined our approach to discover differentiated AI applications in both GenAI and Machine Learning, emphasizing that data access and specificity are common denominators in driving efficiency. We have been adamant that companies must go beyond table stakes features not only to provide value exceeding that of off-the-shelf LLM’s (Large Language Model), but also to future-proof their new investments against the further strengthening of generalized models. CEO of OpenAI, Sam Altman, has similarly warned the ecosystem that few companies will be able to push the leading edge of model development, and that new AI products should approach their deployments assuming that the underlying technology will improve over time. Future model developments will phase out many products whose strengths revolve around small-scale improvements to underlying LLM’s, but the pace at which this will happen is not a linear equation, but rather, a function of the hardware with which models are trained.
As companies push the limits of their hardware footprints to diminishing returns, ROI is brought into the spotlight and non-leading-edge competitors are given opportunity to innovate upon existing LLM’s to disruptive effect. We have seen this first slowdown cycle manifest over the course of the past year, as many hyperscalers who rapidly expanded their hardware footprints have begun to reach the limits of our current generation of chips. This has come to a head in conversations around power expenditure, a theme that we have been following since the initial ramp of hardware investment and seen companies in cooling and power delivery boast great returns as a result. To give some context, a ChatGPT request is nearly 10x more power expensive than a Google search, and AI datacenter power consumption is forecast to quadruple by 2028, a figure that simply isn’t feasible with current energy infrastructure. What this has caused is a slowdown in training results at the leading edge.

According to data from Cohere, return on spend for model development has begun to slow with respect to parameter count per dollar spent, and the compute clusters that continue to train our newest models are seeing tapering results. On the left chart, you can see the trend of maximized hardware footprints limiting the ability to improve upon existing models, and with it, a reduction in return on training spend.
To break this trend and reaccelerate the pace of training, larger or stronger footprints would have to be installed, which will begin again with Nvidia’s new generation of Blackwell systems this year. Alternatively, model training efficiencies could be gained to more effectively use our hardware, but regardless of what drives a reacceleration of training results, what we want to emphasize is that this will be a recurring cycle. New hardware will be released, and the biggest wallets will spend to create bigger and better footprints with which to train. Given the cadence of design for new chips, this will naturally have a cyclical effect where compute limits are reached, and open-source opportunities emerge on each new generation of models as the leading-edge waits for more hardware. We begin to see a recurring cycle, not unlike computing innovation cycles of the past, where first comes the hardware and then follows the software.
As model development slowed over the course of this year, we saw the first examples of hardware limits driving innovation in software. In China, export bans drove the re-architecting of training algorithms and release of DeepSeek R1 V3; the first competitive LLM from the Country built on restricted hardware footprints as a fine-tune of Meta’s LLaMa. Though likely exaggerating performance gains and beginning from an existing LLM as a starting point, DeepSeek was able to develop a competitive product with less hardware through innovations in their training software. In doing so, they’ve improved upon older western LLM’s to reach the same plateau as modern genAI generations and driven significant disruption of valuations for AI-adjacent hardware companies over the course of this week. We have yet to see validity of any potential innovations, as similar training methods have been rumoured to be involved in the development of OpenAI’s o3, but China has shown their ability to circumvent or work within hardware constraints to drive innovations in software. While this is a milestone step for their ecosystem, Chinese domestic products are still behind in the next stage of AI adoption: applicability. For some time now, western LLM’s have largely been commoditized, and the trend of hardware hitting limits of scale has been driving innovation in differentiated applications. This is our second example of hardware constraints driving innovation in software; the commoditization of LLM’s and the creation of AI agents.
As we began to face hurdles in trying to apply bulky, generalised models to much more specific business applications, it became clear to the software landscape that intentional automation of tasks using AI would be much more relevant and monetizable than trying to throw LLM’s at every problem. Agents are much smaller, easy-to-deploy models with access to company or customer data and further training to increase accuracy in completion of a specific task. These are the very aspects of the products we felt would have the most commercial success in our earlier AI commentary, and some of the AI products we have talked about in the past that have been successful in removing humans from their workflow could retro-actively be considered agents. For example, Palantir has been effectively using agentic software in their autonomous defense and intelligence platforms for the last year, as their models are deeply integrated with live data feeds, and autonomously alert threats and assist in strategic decision making given a specific scenario.
Instead of a co-pilot model that simply regurgitates information from the database, Palantir’s defense platforms remove the need for constant satellite monitoring to allow for more efficient flagging of threats and reduced time-to-action for further surveillance. The platform can autonomously suggest deploying an unmanned drone, focusing more satellite feeds on a given area, or give further context to what it has flagged and interpreted as a threat in real time. Beyond their defense applications, they have leveraged their expertise in managing live data feeds and making subsequent intelligent decisions to create an agentic workflow platform for niche businesses in other industries. Through Palantir, companies without traditional AI expertise are able to non-technically build out agents to automate business tasks and realise efficiencies. As a result, Palantir saw significant returns from their AI platforms over the course of 2024 and integrated themselves deeply into the lifeblood of many non-technical businesses. In fact, after the realization of AI gains and the announcement of Palantir’s agentic construction platform last year, the company posted a 340% return in 2024. These first agentic programs will cause large pockets of disruption in many industries due to companies like Palantir, and many software companies will realize lucrative returns as a result, but the technology still has more headroom to improve.
This first generation of agents will become adept at quickly completing simple tasks, but when talking to the SVP of Software Products for IBM and listening to the CFO of OpenAI in December of last year, they both emphasized that this is just the beginning for AI in business development. The key to further innovation lies in the fact that most of the AI we will interact with as a society over the next few years are static thinkers. To give this some context, I’d use Salesforce’s new agentic product rolling out over the first quarters of 2025 as an example. Due to the immense repository of CRM data that Salesforce has access to through their extensive customer base, they have been able to create agentic software that is very good at customer service (CS) workflows. This is much better than their prior attempts at introducing co-pilots, as they have differentiated from table stakes efficiencies and improved their software beyond off-the-shelf LLMs. However, this model is thinking statically, or in other words, it understands very well the typical decision trees that are required to close the most common CS tickets that emerge frequently and in volume for a given business.
Under the hood, the model is given a goal from the outset; close the ticket. Using its vast and complete understanding of every ticket that has been encountered before and the solutions that closed them, it will attempt to accomplish that goal. As you interact with the model, it will begin to eliminate solutions not relevant to your problem and work its way down a static decision tree that leads it to an answer and subsequent action, but its goal will never change. In a dynamic environment, that goal can change, and more intelligent or strategic decisions can be autonomously introduced to improve the result for the business and the customer interacting with the agent. For example, in a given ticket, the model discovers that the customer’s problem is because they have bought a cheaper software package, and that there is an upgrade with more bandwidth and features. The static model would simply explain that they have bought the wrong package, and recommend that they read up on potential upgrades so that it can close the ticket; it accomplished its goal. A dynamic agent would realise that it would actually be better to pass this problem to a sales agent who could explain the benefits of upgrading and why it might be better for the customer than what they have decided to go with. In this scenario, the goal has changed, and from it, an opportunity arises for the business. A static thinking AI dead-set on accomplishing its goal; closing the ticket, would miss this opportunity and likely leave the customer unsatisfied, but a system of dynamic thinking agents could tackle this problem and attempt to see an opportunity. See below a chart that very simply visualizes the difference in decision and results.

What was emphasized in listening to both IBM and OpenAI was that dynamic thinking is the key to agents acting less like automation software, and more like co-workers. OpenAI envisions a world where not only can agents think dynamically, but reflectively, and the goal becomes less about quick answers and efficiency, and more about task offloading. They have laid out concepts of agents that can go off to think about a given problem for a week, or even a month, and come up with a more reflective and accurate result to eliminate tasks completely or provide a starting point from-which human teams can build further. On the other side of the coin, IBM was adamant about the potential for inter-connected webs of agents; where customer service, sales, and client relation programs could be taught to exchange tasks according to relevance and create a front-facing network of AI programs that work in tandem to accomplish strategic incentives. This could be applied to the backend too, as IBM mentions 30% of their code is already being generated by Microsoft’s GitHub Co-pilot. There is a world where programming agents could exchange tasks with product management agents to take easy-fixes and small projects out of the workflow and free up engineers for deeper thought work. In all, agents are the beginning of the wider software landscape thinking smarter, not larger, and from it will come material business applications and the framework of an innovation cycle where software companies can introduce new products and disrupt legacy business development.
Heading into 2025, many of the most applicable AI applications will be agentic technologies. For the first time, non-technical businesses can purchase less-bulky AI products that are specific to their given industry and drive material autonomous efficiencies. As Nvidia rolls out its new generation of chips or LLM developers gain efficiencies from DeepSeek R1, we may begin to see the pace of the leading-edge re-accelerate, but as we know from these past two years, those advancements may not be commercialized as fast as people originally assumed. In those pockets of cyclical innovation slowdown, software companies will be afforded the opportunity to create monetizable AI products like agents to disrupt legacy business architectures and specifically automate tasks for their customers. We feel that the current innovation cycle will be driven by applicability, and that continued differentiation from commoditized LLM’s will be essential to adoption and commercialization. We are of the opinion that this will begin to drive a return on investment, and that there will be opportunities for many companies like Palantir to realise AI gains in 2025 off the back of a year where software companies were largely underappreciated. As always, we will continue to be vigilant as we look for disruptive software companies who are not only well-positioned for this generation of agents, but also future AI technologies that begin to think dynamically and deeply integrate themselves in the strategic process of adopting businesses around the globe.
Disclaimer
This Commentary expresses the views of the author as of the date indicated and such views are subject to change without notice. Wealhouse has no duty or obligation to update the information contained herein. This Commentary is being made available for educational purposes only and should not be used for any other purpose. The information contained herein does not constitute and should not be construed as an offering of advisory services or an offer to sell or solicitation to buy any securities or related financial instruments in any jurisdiction. Certain information contained herein concerning economic trends and performance is based on or derived from information provided by independent third-party sources. Wealhouse believes that the sources from which such information has been obtained are reliable; however, it cannot guarantee the accuracy of such information and has not independently verified the accuracy or completeness of such information or the assumptions on which such information is based.