It's been a remarkable year for AI, not just for the stocks, but because the underlying technology continues to evolve at an incredible rate. Before we address more recent developments and our own outlook, a brief recap might be useful. As you know, generative AI was born in November 2022 following the release of ChatGPT by OpenAI. It was an overnight success, but 66 years in the making. Ever since 1956, when the pursuit of AI began, there had been steady progress as measured by parameter count. Along the way, important breakthroughs too, such as Deep Blue in 1997, IBM Watson in 2011, and more recently the use of reinforcement learning by DeepMind, now Google, to achieve superhuman performance in narrow fields like Go, chess, and predicting protein folding. However, in my opinion, these remarkable achievements look like warm-up acts related to ChatGPT and other large language models that are able to comprehend natural human language and address myriad tasks without task-specific training.
The backbone of large language models, and frankly the new news, is the Transformer, a neural network architecture introduced by Google in 2017. It has literally transformed, apologies, the trajectory of AI progress. These Transformers are able to weigh the importance of different words in a sentence, allowing models to understand context. They're also able to process words all at once rather than one by one or sequentially, making parallel processing possible. As a result, large language models could be trained on enormous datasets. In fact, OpenAI GPT-4, the latest iteration of its AI engine, is said to have been trained on 10 trillion words, a truly ridiculous number that might take a human, if they spend eight hours a day, more than 200,000 years to do. Following the Transformer, the parameter count of models has exploded, now growing 275 times every two years as compared to more like eight times previously.
Now, I've been working as a tech investor for 27 years, believe it or not, and I've heard the term inflection point more times than I care to remember. But I think that this has to be the inflection point to end all inflection points. So much so that we believe that AI not only represents the next general-purpose technology like steam, electricity, and the internet, but also represents a rare example of something called discontinuous technology progress, a sudden and significant leap that can render existing technologies obsolete. For example, the parameter count of OpenAI GPT-4 is rumored to be a million times larger than the model DeepMind developed that beat Lee Sedol at Go just seven years ago. Now, if you go back in history, a good example of discontinuous technology progress, and one of my favorites, was the skyscraper, which transformed the urban landscape thanks to cheap steel made possible by the Bessemer process. Following that, building height increased 25 times in less than 50 years, having previously been relatively unchanged for most of human history.
Now, to those alive in 1885, this transformation wouldn't have been obvious, not just because humans struggle with nonlinear change, but particularly when it's discontinuous. And early iterations of general-purpose technologies are often imperfect, even underwhelming. If you had been there in 1885 looking at the Home Insurance Building in Chicago, the first steel-frame structure ever, you may have been underwhelmed. It was, after all, just 10 stories, and it looked a little like existing structures. The same feels true today. Early AI applications look a bit like existing products. AI chatbots that we use resemble earlier chatbots. AI search looks a bit like using Google. Software copilots that are today transforming productivity still are helping existing applications. However, as with skyscrapers, our view is that this current state is unlikely to persist for very long. And in fact, we think about the Transformer as the Bessemer moment for AI, a breakthrough that we expect to radically and rapidly transform the technology landscape, which is why we significantly pivoted our portfolios towards AI last year and introduced an AI lens to our investment process.
Now, our AI excitement begins with the fact that the market opportunity is vast, given that AI in the end addresses knowledge work. McKinsey believed that generative AI might automate 30 to 50% of tasks across 60% of occupations by 2030. Longer term, we believe the opportunity might be greater still, should AI begin to substitute rather than augment human labor. For instance, today technology budgets account for about 3 to 6% of corporate revenues, but as AI can begin to substitute for human capital, so AI may address the overall corporate OpEx part of the P&L, a much larger opportunity. But differently, this year around $5 trillion will be spent on it, as compared to more than $40 trillion expected to be spent on wages at the margin. And incredibly, a little less than just two years after ChatGPT was launched, we are already beginning to see this substitution effect in the contact center, today a $23 billion software market supporting 7 million human agents. However, as AI begins to replace humans, the addressable market might expand towards $500 billion, said to be the fully loaded cost of human agents.
By now, buy now pay later provider Klarna has been particularly vocal about their use of AI, which today they are said to be using to replace 700 full-time employees, saving the company $40 million per annum. While data points such as these inevitably dominate headlines, AI today remains focused on making humans more productive or helping us with tasks that we struggle with. Walmart recently commented on how they use large language models to create or improve more than 850 million pieces of data in its product catalog, which without generative AI would have required nearly 100 times the current headcount to complete in the same amount of time. Tools that enable knowledge workers to be more productive, so-called copilots, are already making exciting progress and enjoying strong adoption. Microsoft's GitHub Copilot already helps software developers write 46% of code in half the time, while Axon's DraftOne, a newly released piece of software from the company that makes tasers, is able to auto-draft police reports using body cam footage.
Now, writing reports can account for 30 to 40% of US police time and is said to be the number one reason for early retirement. DraftOne can meaningfully reduce this by producing 80% of a police report in just six minutes, reducing time spent on repetitive tasks. Making knowledge workers more productive is a massive opportunity, potentially worth hundreds of billions of dollars, reflecting the fact that there are one billion knowledge workers in the world today. Longer term, we expect AI to create myriad new and large markets in areas like drug discovery, autonomous vehicles, and humanoid robots, together with many that we cannot know today. Despite this and today's opportunities, some people remain myopically focused on identifying the killer app that will drive widespread AI adoption. Obviously, we would love to know the answer to this question too. In fact, it's one of the questions we ask our companies every day. But we are comforted by the fact that history says it's almost impossible to know this early in a new cycle.
For instance, the app economy that today is worth trillions didn't exist in 2007 when the iPhone was first introduced. Even when there have been compelling early visible market opportunities, these are later eclipsed often by larger invisible opportunities at the time. For instance, in steel, after the Bessemer process, railways accounted for around two-thirds of steel because steel rails were vastly superior to iron rails. But by the mid-20th century, rails represented just 15% of the steel market as applications that didn't exist in 1856, such as automotive, drove incremental demand. More recently, you might remember how early connected devices such as digital cameras and music players, including the trailblazing iPod, were later subsumed by the smartphone market. We are also encouraged by the furious pace of AI model improvement that continues to make the technology ever more capable.
This year alone, we've seen OpenAI demo Sora, a remarkable text-to-video model. We've seen embodied AI in the humanoid robotic form of Figure AI, as well as Meta's latest 420 billion parameter open-source large language model, Llama 3. While GPT-5 is still expected before year end, OpenAI also recently previewed a new model called o1, which applies something called chain-of-thought reasoning, the use of sequential intermediate steps to improve accuracy. One exciting white paper suggests that with enough steps, chain-of-thought models such as o1 might be capable of solving any solvable problem. With more data, compute, and parameters, scaling laws also continue to predict future model improvement, albeit at a growing cost. Models are trending towards 10 trillion parameters, which means that GPT-6, should it happen, could cost between $25 and $30 billion.
Recently, the CEO of Oracle, Larry Ellison, stated that within three years he thought frontier large language models might cost as much as $100 billion. Finally, we're comforted by the fact that some of the world's smartest technologists and business leaders are already vying for AI leadership, as are nation states that are expected this year to spend more than $10 billion with Nvidia alone. The prize: a mix of defending existing market dominance, leadership of new enormous market opportunities like autonomous vehicles, or perhaps even the greatest prize of all, the one that few dare mention, artificial general intelligence, the point where AI meets and then ultimately surpasses human levels of cognition. The race for AI leadership has led to an explosion in AI-related spending, in part because existing cloud infrastructure was not designed for parallel processing or accelerating computing.
As such, we believe AI represents a break from general-purpose CPU-based compute towards a new form of accelerated computing based on GPUs and custom ASICs, reflected by a 25 times higher bill of materials of an AI server when compared to a traditional one. We liken this contrast to a Formula 1 car and a Toyota Prius, where every component in an AI server is designed for performance rather than cost in order to viably train a large language model on vast datasets. This year, US cloud company capital expenditure on AI is expected to reach more than $150 billion, equivalent to around half a percent of GDP, leading some to argue that we're in the midst of a capex bubble. We disagree strongly. While growth has been truly remarkable, we believe we're still early in what we think is an accelerated AI infrastructure build. By 2027, expectations are that spending could surpass $1 trillion, equivalent to what has been spent to date building out cloud infrastructure.
Now, this embeds certain assumptions around user growth and daily usage growth, but these don't look unreasonable given the internet experience and also the fact that we spend four and a half hours a day on our smartphones, which immediately suggests there's plenty of upside to the current three minutes of daily usage in inference. Moreover, half a percent of GDP after just 18 months of investment looks early relative to historic infrastructure booms. While we all know the 19th century railway boom ended in a bust, it wasn't before investment grew from 0.2% of US GDP to just over 0.9% between 1830 and 1839. Actually, that represented a 31% compound growth rate, not too dissimilar to the 27% that we expect between 2022 and 2026 in AI. And after a digestion period, railway investment then averaged 1.7% of GDP for almost a decade, reaching 2.6% in 1854.
AI spending today, at just 0.54% of GDP, and in just year two, we believe that AI spending remains far from bubble territory. With naysayers focused on an apparent lack of killer app, we believe that progress towards AGI, artificial general intelligence, and the ability to answer almost any question, solve any solvable problem, will drive exponential demand for AI. In our bull case outcome, the volume explosion we envisage could resemble the 65-fold increase in digital messages that we send today compared to the number of letters that we once sent, or the 40-fold increase in time spent video gaming compared to gaming before the internet and smartphones. While the internet has reshaped how we access information, AI has the potential to reimagine the relationship between humans and machines, a little like we saw in the agricultural and industrial revolutions.
Like the telegraph, we think AI will transform the speed of knowledge creation, improve R&D productivity, and act as an agency for the alteration of ideas. As such, AI has the potential, in our opinion, to reduce the internet to something like AI source material and a dumb pipe, in much the same way that the internet did to Yellow Pages, libraries, and communication networks before it. We truly believe the AI story is just beginning. Where others may predict steady diffusion, we expect AI adoption and related disruption to follow electrification, which was sweeping and widespread. For now, we're focusing our portfolios on the companies helping to build the AI rails, the most direct beneficiaries of an infrastructure build that remains in our view. While we expect elevated levels of volatility to persist, normal early in a new technology market, we remain hugely excited about the vast and unbounded potential of AI, which is likely to prove, in our opinion, the most important theme of our investment lifetimes.
