Summers and winters in artificial intelligence and virtual reality, traced through the events that bent each curve — beginning with McCulloch & Pitts in 1943 — the breakthroughs, the bankruptcies, the books, and the buyouts.
Before AI had a name, it had a mathematics. McCulloch & Pitts (1943) proved that networks of binary threshold neurons could in principle compute any logical function — the seed of every later neural model from Rosenblatt’s Perceptron through AlexNet to Transformers. Hebb (1949) supplied the first credible learning rule. Turing (1950) framed the inheritance question. By the time McCarthy convened Dartmouth in 1956, the conceptual scaffolding already existed; what Dartmouth added was a name, a programme, and an organised funding pitch.
Three summers, two real winters. The first AI summer was a funding-and-promise bubble around symbolic reasoning that collapsed under combinatorial explosion and overpromising — Lighthill in 1973 and DARPA cutting speech-understanding in 1974 closed the era. The second was an industrial bubble around expert systems and Japan’s Fifth-Generation programme, killed when LISP-machine vendors were obsoleted by general-purpose Sun and Macintosh workstations after 1987.
The interregnum mattered. The 1993–2011 “quiet” years are sometimes mis-remembered as a winter, but capability advanced steadily under disguised names — machine learning, statistical methods, kernel methods, data mining. Hinton’s 2006 deep-belief-network paper rebranded neural nets as “deep learning,” and ImageNet (2009) gave the field a benchmark to chase.
The current summer started with hardware. AlexNet (2012) was not algorithmically novel — it was CNNs running on GPUs, the same McCulloch-Pitts threshold neurons in a deeper stack. The Transformer (2017) then provided the architecture that scales with compute, and ChatGPT (Nov 2022) compressed eight decades of research into a consumer product moment. Whether the current cycle ends in a winter, a plateau, or general AI remains open. Confidence: high on history, low on forward call.
VR’s cycles run on hardware physics, not algorithms. The 1985–95 summer collapsed because the silicon could not deliver what the marketing promised — the Virtual Boy is the canonical artefact, but the deeper failure was that PC graphics in 1995 simply could not render an immersive scene at sufficient resolution and framerate without inducing nausea. Sixteen years of Moore’s Law had to pass before Oculus could reasonably try again.
The 2014 Facebook acquisition was a 7-year head fake. Zuckerberg’s $2 B Oculus bet bought the company a generation of capital, but consumer VR underperformed every analyst forecast between 2016 and 2019. Quest (2019) and Quest 2 (2020) finally cracked the price-and-comfort curve. Then COVID, idle attention, and crypto-fuelled liquidity converged into the metaverse hype peak of late 2021.
The metaverse winter was largely an AI eclipse. Cumulative Reality Labs losses crossed $46 B by 2023. ChatGPT (Nov 2022) didn’t just compete for headlines — it competed for engineering talent, capital allocation, and executive attention at every large tech company. Apple’s “spatial computing” rebrand in 2024 is the same hardware story under a new label; Vision Pro production was cut back by 2025. Confidence: high on numbers, moderate on causal weighting of the AI eclipse — it is one of several factors.
AI has a twenty-five-year head start. The combined chart now extends to 1943, and the asymmetry is hard to miss — McCulloch & Pitts published the first computational neuron the same year ENIAC was being built, while VR’s lane is empty until Sutherland’s head-mounted display in 1968 and only really populates from 1984 onward. AI is fundamentally an older field with a deeper bench of theoretical results. VR has always been catching up to a moving target.
Read the combined chart vertically. The 1990s are the only true shared winter — expert systems collapsed and Virtual Boy flopped within two years of each other. Both fields then disappeared from headlines for roughly fifteen years while their underlying technology quietly kept improving (machine learning rebrand on one side; commodity 3D acceleration, IMUs, smartphone display panels on the other).
2012 was the shared ignition year. AlexNet and the Oculus Kickstarter are within a few months of each other. This is not coincidence — both depended on commodity GPUs becoming powerful enough to do something previously confined to specialised hardware. Two years later, in 2014, Facebook bought Oculus and Google bought DeepMind. The same firms made parallel bets on both fields.
The 2022–24 anti-correlation is the most consequential dynamic of the decade for anyone building in either space. ChatGPT arriving thirteen months after Facebook’s Meta rebrand redirected almost every relevant capital pool — venture, talent, executive attention, GPU allocation — toward LLMs and away from headsets. The metaverse winter is real but only partially endogenous; a large component is exogenous, an AI eclipse. Confidence: high on the timing, moderate on the causal magnitude.
AI cycles are funding-and-credibility cycles around algorithmic claims. VR cycles are hardware-and-adoption cycles around display, optics, and silicon. Treating them as the same kind of bubble is an analytical error.
Winters are usually triggered by a small number of nameable events: a damning report (Lighthill), a dismantling book (Perceptrons), a cancelled product (Virtual Boy), or a vendor collapse (LISP machines). They are not slow declines — they are step functions.
Hype is a near-zero-sum resource. When ChatGPT arrived in late 2022, it didn’t merely outshine the metaverse — it cannibalised the capital, GPUs, and engineering attention that the metaverse pitch had just claimed. Spatial computing’s 2024 rebrand is in part an attempt to escape that gravity well.