The Rapid Evolution of AI and Why Its Acceleration Is the Strategy

The Long Runway: 1958 to 2012

The ideas underneath modern AI are old. In 1958, Frank Rosenblatt described the perceptron, a model that adjusted connection weights to learn a classification from labeled examples, and built it into hardware as the Mark I Perceptron at the Cornell Aeronautical Laboratory [1]. The press treated it as the dawn of thinking machines. The reckoning came in 1969, when Marvin Minsky and Seymour Papert proved that a single-layer perceptron could not represent functions as basic as exclusive-or [2]. Funding and attention drained away, and the field entered the first of two long stalls later called AI winters.

The algorithmic answer to that limitation arrived in 1986. David Rumelhart, Geoffrey Hinton, and Ronald Williams showed that backpropagation could train networks with hidden layers, letting internal units learn useful features instead of having them hand-designed [3]. Yann LeCun and colleagues applied the method to convolutional networks for handwritten-digit recognition by the end of that decade, and Sepp Hochreiter and Jürgen Schmidhuber added long short-term memory in 1997 to handle sequences [4]. The methods worked on small problems and stayed academic. Through the 1990s and 2000s, neural networks lost ground to support-vector machines and other approaches that performed better on the modest datasets of the era. The algorithms that now run the field existed for decades before the field could make them pay.

What changed in the last decade was not a new idea. It was the arrival of the two inputs those ideas had always needed: data at scale, and compute cheap enough to use it. The internet supplied the first. No laboratory could hand-label enough images to train a large network, but the web already held billions of them, and in 2009 Jia Deng, Fei-Fei Li, and collaborators organized millions into the labeled ImageNet hierarchy [5]. NVIDIA supplied the second. Its CUDA platform, released in 2007, made graphics processors programmable for general parallel mathematics in C, turning commodity gaming cards into affordable matrix-multiplication engines [6], [7]. Backpropagation through a deep network is mostly matrix multiplication. Once a large labeled corpus and a programmable parallel processor existed in the same few years, the dormant methods finally had what they had been missing.

The lesson for the strategy in this paper is in how those inputs combined, not in any one of them. Data and compute, not a fresh theory, are what turned a fifty-year-old idea into a working system. Both keep compounding. That is why the intervals after 2012 keep shrinking.

2012: AlexNet and the GPU Inflection

In the 2012 ImageNet Large Scale Visual Recognition Challenge, a University of Toronto entry built on exactly those inputs won decisively. AlexNet finished with a top-5 error rate of 15.3 percent against 26.2 percent for the second-place system, a margin of 10.9 points in a contest where the runner-up sat near the previous year’s winning score [8]. The gap marked a break, not an increment, and the field reorganized around it inside roughly two years.

The hardware mattered as much as the result. Alex Krizhevsky trained the 60-million-parameter network on two NVIDIA GTX 580 gaming cards with 3 GB of memory each, and the paper credits feasibility directly to “a very efficient GPU implementation of the convolution operation” [8]. This is the empirical origin of the GPU-as-AI-accelerator thesis that now underwrites NVIDIA’s data-center business and every hardware decision in this paper. A research result enabled by commodity hardware became the foundation of a hardware giant’s roadmap within a decade. The path was visible in 2012. Most organizations did not act on it until 2023.

2017: The Transformer

Five years later, eight researchers, most of them at Google Brain and Google Research, published “Attention Is All You Need.” The architecture dispensed “with recurrence and convolutions entirely,” discarding the recurrent networks and LSTMs that had defined language processing for a decade [9]. On the WMT 2014 English-to-German benchmark it reached 28.4 BLEU, more than two points above the prior best; the larger configuration set a single-model English-to-French record of 41.8 BLEU [9].

The benchmark numbers undersell the point. The Transformer’s real contribution is architectural: it scales predictably, where adding parameters, data, or compute yields capability gains along smooth curves that recurrent models never offered, a relationship later formalized as neural scaling laws [10]. Every production language model discussed in this paper traces to it. The base Transformer trained in about twelve hours on eight NVIDIA P100 GPUs [9]. From the start, training large transformer models has required multiple GPUs, and the scale of that requirement has only grown. Epoch AI measures frontier training compute growing four to five times per year, which compounds to several orders of magnitude across the decade [11].

2020: GPT-3 and Few-Shot Capability

In a May 2020 paper, OpenAI described GPT-3, a 175-billion-parameter language model an order of magnitude larger than any previous dense model [12]. The technical result was a model that performed tasks it had never been trained on, from a natural-language instruction and a handful of examples, “without any gradient updates or fine-tuning” [12].

This is the empirical basis of the foundation-model idea, which a Stanford group named the following year: one large general-purpose model, trained once, applied across translation, question-answering, code generation, and reasoning without retraining for each task [13]. Capability stopped being built task by task. It is built once, at scale, then specialized.

November 2022: ChatGPT and the Board-Level Inflection

ChatGPT launched on November 30, 2022. Five days later Sam Altman, OpenAI’s chief executive, posted that it had crossed one million users [14]. By UBS’s estimate, drawn from SimilarWeb traffic data and reported by Reuters, the service reached roughly 100 million monthly active users about two months after launch, a faster consumer-app ramp than the analysts could recall in twenty years of covering the sector; for comparison they cited about nine months for TikTok and two and a half years for Instagram [15]. That figure is an outside estimate rather than an audited count, but its order of magnitude held up.

ChatGPT introduced no new architecture. GPT-3.5 was an iteration on the 2020 work. What changed was delivery: a chat interface put transformer capability in front of non-technical users at a scale the field had not pictured.

This is the moment AI moved from research agenda to board agenda. Executives who had ignored a decade of deep-learning papers asked their staff for a position on generative AI within ninety days. Organizations with an answer were positioned; the rest began a catch-up cycle many are still inside.

2023–2024: Open Weights and the Self-Hosting Precondition

In July 2023, Meta released Llama 2 with Microsoft, free for research and commercial use [16]. This is the operative event for the strategy in this paper. Llama 1, released five months earlier, had been restricted to academic researchers granted access case by case [17]. Llama 2 removed that restriction and triggered the open-weight wave that now includes Llama 3 and 4, Qwen, DeepSeek, Mistral, Gemma, Nemotron, and many others.

These models are open-weight, not open-source. Meta’s license carries acceptable-use and large-scale commercial conditions that fall outside the Open Source Initiative’s Open Source Definition, which is why the OSI rejects the “open source” label for them [18]. What an organization may legally do with a fine-tuned derivative depends on which side of that line its base model sits.

Before July 2023, an organization had two options: cloud APIs that moved its data off-premises, or research models it could not deploy commercially. Llama 2 created a third: self-hosting a commercially licensed model, fine-tuned on proprietary data, inside the organization’s own boundary. Every roadmap recommendation in this paper depends on that enabling condition.

2024–2026: AI as Agent

The current inflection is the shift from AI-as-assistant to AI-as-agent. An assistant suggests and the human accepts or rejects. An agent plans a multi-step goal, executes it through tool calls, checks the result, and continues without sign-off at each step until the goal is achieved.

The 2025 wave defined the transition. Anthropic shipped Claude Code with Claude 3.7 Sonnet in February 2025; OpenAI launched the Codex cloud software-engineering agent in May 2025; GitHub’s Copilot coding agent reached general availability later that year. Anthropic’s documentation describes Claude Code as running an “agentic loop” in which the model gathers context, takes action through tools, and verifies the result before repeating [19]. The previous generation suggested code. This one writes it, runs it, reads the errors, and iterates.

The adoption data is the harder evidence. Anthropic’s February 2026 analysis of Claude Code and public-API traffic found that full auto-approve, where the agent acts without per-step human sign-off, rose from about 20 percent of sessions among new users to over 40 percent as users gained experience; an independent analysis of the same release places the 40-percent crossover near a user’s 750th session [20], [21]. Over the same window the 99.9th-percentile session length nearly doubled, from under 25 minutes to over 45 minutes, and that rise was smooth across model releases rather than stepwise with each launch [20]. Anthropic reads the smoothness as a deployment overhang: the models were already capable of more autonomy than users were granting them, and trust, not capability, was the limiting factor.

The closed-source frontier has moved faster than any annual planning cycle can absorb. Anthropic shipped Claude Opus 4.7 on April 16, 2026, then replaced it with Opus 4.8 on May 28, the shortest gap yet between Opus releases [22], [23]. On June 9 the company opened a Mythos-class tier above Opus with two models: Fable 5, generally available with safeguards, and Mythos 5, restricted to the invitation-only Project Glasswing program for defensive cybersecurity [24]. Three days later, on June 12, a US government export-control directive forced Anthropic to disable both Fable 5 and Mythos 5 for every customer worldwide; Opus 4.8 and the rest of the line stayed available [25]. The Department of Commerce lifted the controls on June 30, Anthropic restored Fable 5 globally on July 1, and Mythos 5 returned only to a set of approved US organizations, a 19-day suspension start to finish [26]. The most capable models a vendor offered went dark worldwide, with no notice, at the direction of a party outside the customer relationship, and returned on changed terms. That is the availability risk this paper treats as central, made concrete.

Two clarifications on how this paper treats the frontier. First, the technical respect is real: as of mid-2026, the Anthropic, OpenAI, and Google frontier models lead the best open-weight options on many agentic and reasoning benchmarks. Second, that respect is separate from the API-only delivery these vendors require. The paper’s aim is to identify open-weight models that approximate frontier capability while running inside an organization’s own perimeter. Which model is best, and which model an organization can responsibly operate on its own data, are different questions with different answers. The Fable 5 suspension is the clearest recent demonstration of why on-premises operation matters.

2026 Onward: Smaller Models, New Architectures, and the Move to Local Compute

The trajectory through 2026 runs along two axes at once with both favoring operation inside an organization’s own boundary. The first is compression. NVIDIA researchers argue that small language models, which they treat as models compact enough to run on consumer or single-GPU hardware, are already capable enough for the repetitive, narrow calls that make up most of an agentic workflow, more suitable for those calls than a general model, and the economically correct default for them; the practical design they advocate is heterogeneous, with a large model reserved for the hard reasoning steps and fit-for-purpose small models handling the routine ones [27]. The consequence for this paper’s argument is direct: the work an agent does thousands of times per task is precisely the work that no longer needs a frontier model or a data center, and the open-weight wave described earlier now supplies enough capable sub-10-billion-parameter models to staff that role on hardware a firm can own.

The second axis is architectural, which shows that the capability curve is broadening rather than flattening. The autoregressive transformer that carried the field from 2017 is no longer the only serious bet at the frontier. In June 2025 Meta released V-JEPA 2, a 1.2-billion-parameter video model built on the Joint-Embedding Predictive Architecture that Yann LeCun’s group first proposed in 2022; it learns a predictive model of physical dynamics in a latent representation space instead of generating pixels, and after fine-tuning on 62 hours of robot interaction data it supports zero-shot robot planning in unfamiliar environments [28]. LeCun’s position is that predicting in representation space, not pixel space, is the route to systems that plan and act in the physical world. Whether JEPA-style world models displace or complement autoregressive language models is unsettled, but a well-funded architectural alternative at the frontier is itself evidence against an imminent plateau.

The same turn toward the physical world appears in the rise of physical AI as a distinct model class. NVIDIA’s Cosmos world foundation models, introduced in 2025 and extended with the open Cosmos 3 omnimodel in 2026, generate physics-grounded synthetic experience to train robots and autonomous vehicles, easing the real-world data-collection bottleneck that has held robotics back [29], [30]. Jensen Huang’s framing, that world foundation models are to physical AI what large language models were to generative and agentic AI, is a vendor’s claim and should be read as one, though the adoption by robotics developers building on the models is real [30]. For a firm planning AI infrastructure, physical AI matters less as a near-term purchase than as a signal: the capability frontier is widening from text and code into embodied domains, which lengthens the runway over which today’s infrastructure investment compounds.

Underneath both axes, the semiconductor industry is provisioning for sustained demand rather than a plateau. TSMC began volume production of its 2-nanometer N2 node, its first generation built on gate-all-around nanosheet transistors, in the fourth quarter of 2025 [31]. At its 2026 technology symposium the foundry projected its 2nm capacity to grow at a 70 percent compound annual rate from 2026 through 2028, with advanced-packaging capacity, the CoWoS and SoIC processes that bond logic dies to high-bandwidth memory, expanding more than 80 percent per year through 2027 [32]. A foundry does not commit that capital against a demand curve it expects to soften. The process advances that feed data-center accelerators also raise what an organization can run on a single workstation or a small on-premises cluster. That is the trend that turns the on-premises case from an efficiency argument into a capability one: each node generation narrows the gap between what a firm can run inside its boundary and what it must rent outside it.

The Acceleration Argument

The intervals are not merely short. They are still contracting. Five years from AlexNet to the Transformer. Three from the Transformer to GPT-3. Two and a half from GPT-3 to ChatGPT. The agentic transition went from research demos to production deployment in roughly eighteen months. The pace shows up in direct measurement: METR finds that the length of task a frontier agent can complete autonomously at 50 percent reliability has doubled about every seven months since 2019, with a shorter doubling time on software-specific tasks, a trend whose task selection and extrapolation remain contested but whose direction is not [33]. Frontier training compute grows four to five times per year over the same period [11]. Annual hardware-planning cycles are slower than the thing they plan for.

This pace is why a phased roadmap should be read as a capability-building program, not a fixed technology plan. The specific GPUs in a Phase 1 recommendation will be superseded within thirty-six months of purchase. What compounds across silicon generations is everything else: the workflows built on the first hardware, the models fine-tuned to internal data, the staff who learned to run the stack, and the institutional fluency with prompt design, evaluation, and deployment that no later purchase restores.

The implication for an organization still on the sidelines is specific. The on-ramp does not get easier with time. Models improve, but competitor workflows improve faster, vendor pricing power consolidates, and the learning curve steepens as the systems grow more capable. A firm starting a Phase 1 pilot in 2026 sits in manageable catch-up territory. The same firm starting the same pilot in 2029 faces competitors with over three years of validated workflows, models trained on their own proprietary corpora, and staff who already absorbed the move from assistant to agent. The Fable 5 suspension is the other half of the warning: capability rented through an API can vanish on a Friday afternoon at a third party’s direction and return, weeks later, on changed terms; capability built inside the boundary does not. Neither the head start nor the independence is purchasable after the fact, at any price.