Revenge of the Nerds |
A lot of people drop “AI” into their equity story these days. For pure play AI we get most excited about those stories that involve privileged access to at least one of three capabilities: chips, electrons or better maths. On the last capability, our view is that the AI race has only just begun to appreciate the importance of a specialised data science and maths layer that introduces things like proprietary weights, data ontologies, model interoperability and composability, the open world, and, last but certainly not least, new mathematical methods. Commodity LLMs sit on top of generally trained, tokenised data, and therefore will always understand the world in a limited way – in a static, Bayesian distribution kind of way[1]. With 50% of internet traffic (and growing) now being bot-driven the overall data quality and signal to noise ratio that underpins these models may diminish. We think that in another 6-12 months’ time the large language models will start to only eke out marginal gains of function. Whereas a lot of people think the models are becoming “so good” that even maths is now commoditised, we don’t. Our thesis is that there’s going to be a Revenge of the Nerds, coming soon. New ground in AI will not be captured, in our view, by the existing broad application and distribution layers (think Google or Microsoft or Apple) but rather through new maths and ontology customisation (an ontology is a map that tells an AI what a business’s data actually means and how it interacts). For us, this will be the “picks and shovels” play. What makes the past few months so instructive is that the main players in the capex-intensive AI infrastructure layers (the chips and the electrons) are also behaving like this. |
Nobody at the top of this race believes their current distribution moat is durable, and everyone is buying the adjacent layers (electrons, chips, maths) while their equity is expensive.
Chips |
Computer games evolved from an Atari pixelation into machines painting enormous grids of numbers in parallel. GPUs render these pixels by applying the same, simple arithmetic on millions of points at once. This is exactly what a neural network model needs. Training and running a model with matrix multiplication repeated billions of times. It’s called “vector compute”.[2] Nvidia’s genius was recognising this early and building a software layer called “CUDA”, that let researchers play with a graphics card as a new kind of general purpose vector machine. And with that masterstroke, a new moat was born. But wait, there’s more. Since every token shuttles model weights between memory and processor, bandwidth and capacity (not raw compute) becomes AI’s rate limiter. Dynamic Random Access Memory (DRAM) operates as the short term scratchpad that holds whatever the processor is actively using. How much memory a chip has, and how quickly it can be read, decides how much useful work the chip does. A model’s weights sit in the DRAM while the model runs. We estimate DRAM chips will absorb 30 to 40 per cent of hyperscaler capex next year, and the premium variety (memory stacked in towers beside the processor, known as high bandwidth memory (HBM)) can only be made by three listed companies: SK Hynix, Samsung and Micron.[3] |
Electrons |
Power, not demand, is also a rate limiting factor on the roll-out of the current AI models. Only the hyperscalers and a select few (like CoreWeave) are bringing on gigawatts of capacity. People are searching for solutions. Nuclear will take a lot more time than people, acclimatised to software, currently imagine. Tesla supercharger sites might be useful for modular data centre units on the edge, but that won’t be scalable. Gas power will be limited by turbines and approvals. Meanwhile, Elon is looking at the stars. We will write about data centres in space soon (hint: we’ve written about how bullish we are on launch cadence and the amount of metal that is going to be sent up to space). Even the Frontier Models are seeking to use electrons from a rocket company. The luckiest equity here has been the bitcoin miners. They spent a decade doing the unglamorous work of securing grid connections, building substations, locking in cheap power – to feed a different chip and a different business model. Their economics collapsed in 2024. Australian-founded IREN converted its Texas campus into an AI cloud anchored by a Microsoft contract worth roughly $1.9 billion a year at an estimated 85 per cent EBITDA margin. Rarely has an industry been so comprehensively rescued by a disruptor technology. Since when do disruptors look out for each other? |
Nvidia buys its way into maths |
Nvidia’s chip position remains extraordinary, yet its recent transactions read like a company preparing for the day it is not. In December 2025 it acquired key technology and personnel from inference specialist Groq in a $20 billion arrangement. In February 2026 it bought Illumex, a data semantics business whose entire purpose is mapping the meaning of enterprise data. In June it added Kumo AI for upwards of $400 million, picking up foundation models that predict business outcomes from a company’s own relational data. Alongside these sit a $2 billion stake in engineering software house Synopsys and participation in TerraPower’s $650 million nuclear raise. And just announced, a partnership with Palantir to deliver “sovereign AI” for US government agencies and critical infrastructure. Nvidia’s Nemotron models are being folded into Palantir’s data science products, to allow agencies to post-train frontier class models on their own data inside classified, air gapped environments and, critically, retain full ownership of the resulting model weights. This is the maths leg in exactly the broad sense we mean it: the value sits not in any single model but in the marriage of proprietary weights, the ontology that maps what an organisation’s data actually means, and the pipes between them. When the world’s dominant chipmaker needs Palantir’s semantic layer to land its models in the buildings that matter, the maths is not an accessory to the stack – it is the point of control. |
The coming flood of open-source models |
The model layer is also commoditising in real time. Chinese open-weight models – DeepSeek, Alibaba’s Qwen, Z.ai’s GLM – now account for roughly 30 per cent of global model usage, and enterprises are already routing coding and cost-sensitive workloads to them at roughly a sixth of frontier model pricing. Because the weights are free to download and can run on a company’s own servers, even the data-sovereignty objection is being mitigated. |
In a nutshell |
| When a “lookalike” frontier model costs next to nothing, nobody will pay a premium for the LLM layer. What is valuable is the new maths and new data science – a map of what an enterprise’s data actually means as it unfolds in real world contexts. Enterprise adoption of open-source models does not threaten our thesis; it accelerates it. |
[1] LLMs treats a fact as a tokenised, fixed, context-free unit. A token’s meaning is not something embedded in a situation, so its meaning cannot shift as the circumstances around it change. A deduction that holds in one situation can be flat wrong in another, and handling that shift is the hard part. Potentially proprietary weights can deliver some ‘error correction” but those weights are still not completely dynamic or fully contextual. [2] A CPU is a handful of very clever cores working through tasks one after another; a GPU is thousands of simple ones doing the same sum simultaneously. [3] A US start-up called “Positron” is the countertrade. The US startup builds inference only accelerators that sidestep the HBM crunch entirely, using commodity low power memory (currently used in phones and laptops) to put more than 2 terabytes on its forthcoming Asimov chip against roughly 384 gigabytes for Nvidia’s Rubin, while claiming five times the tokens per watt. A $230 million Series B in February valued the company above $1 billion. The pitch is not to beat Nvidia at training but to co-exist on the leftover power: air cooled racks drawing about 3 kilowatts that slot in where the grid, not the GPU, is the constraint. OpenAI, meanwhile, has started making its own chip with Broadcom – the maths layer reaching down into silicon just as Nvidia reaches up into maths. |
|
Disclaimer: This blogpot is for informational purposes only, it does not constitute investment or financial advice. Nothing here constitutes an investment offering of any kind, unless expressly stated. Before acting on any information contained in this blogpost, each person should obtain independent taxation, financial and legal advice relating to this information and consider it carefully before making any decision or recommendation. The Powerhouse Group of Companies holds an Australian financial services licence (AFSL# 497505) and may have a shareholding in some companies discussed. Visit our website to read our privacy policy. |