Is AI a hype cycle or a genuine technology shift?

Executive overview

AI investment is concentrated and fast-moving, raising legitimate questions about whether valuations are outpacing reality. The useful distinction is between short-term price inflation — which may be real — and long-term value creation, which the evidence increasingly supports.

The voting machine vs. weighing machine framework cuts through the noise: in the short run, markets reward credentials, hype, and social momentum; in the long run, only retained customers and cash flows matter. Early AI companies are already showing the latter.

What makes this moment different from past hype cycles

• Both startup and public markets are simultaneously AI-driven — an unusual degree of synchronisation
• The "magnificent seven" stock gains are almost entirely AI-attributed, the most concentrated market leadership in history
• Unlike crypto, AI products pass an immediate sniff test: summarising a 50-page report or automating accounts receivable has obvious, measurable utility
• The ChatGPT-wrapper fear from early 2023 has been largely debunked — multiple competitive models now exist, including open-source at near-frontier quality
• Llama 3 reaching near-parity with frontier closed models was not anticipated even a month before release

Where value actually accrues

• The stack has four layers: chip makers, hosting providers, foundation model developers, application companies
• Application layer companies need no capital-intensive infrastructure — a laptop and an internet connection suffice
• DoorDash and Instacart are the template: application-layer winners built on someone else's infrastructure shift
• It took ~4 years post-iPhone for the defining mobile application companies to emerge; AI timelines may follow a similar pattern
• Even if all model progress froze today, Zuckerberg's estimate is 5 more years of application-layer innovation remain
• GitHub Copilot reportedly accounts for 40% of GitHub's recent revenue growth, illustrating how fast application-layer value can compound

The open-source and model competition shift

• Six to twelve months ago, ~80–90% of YC batch companies used OpenAI models exclusively
• A recent casual survey shows significant share shifting to Claude 3.5 and Llama — OpenAI usage declining as models become competitive
• Open-source models were previously 6–12 months behind frontier; they are now at near-parity
• Commoditisation of models is a feature for application-layer builders, not a threat

Signs of real traction at the application layer

• YC batch revenue grew from $6M aggregate at application to $20M over 3–4 months — exceeding the 20% month-over-month growth benchmark
• One legal AI company (Leo) closed a Series A from Benchmark; another is on track for $10M ARR one year after finding product-market fit
• Accounts receivable automation: a 12-person team reduced to 1 person on that function
• Call centre replacement: hundreds of thousands of calls processed, offshore BPO centres shut down, at 20–100x lower cost
• PhotoRoom (generative image AI for e-commerce) reached a $500M valuation on real revenue
• PermitFlow (construction permit automation): winning through sales execution and UI detail, not model differentiation
• A YC company recently discovered an entire industry — unknown to most technologists — with no AI competition and a clear billion-dollar fit

The overvaluation question — and why it matters less for founders

• Some AI valuations are likely inflated (e.g. Nvidia at peak, early-stage teams with zero revenue and $500M raised)
• The crypto parallel: "professor coin" dynamics — credentialed teams raising at billion-dollar valuations without a line of code
• Key difference: AI companies with real revenue are hitting profitability early and not needing follow-on rounds
• Zapier (only ever raised a seed round, now hundreds of millions in ARR) and Weebly are precedents for capital-light compounding
• For founders and early-stage investors with 10-year horizons, short-term overvaluation is largely irrelevant — and may even accelerate ecosystem development
• The dangerous scenario is large raises with zero revenue: a $200–500M balance sheet with no path to monetisation is a different risk class entirely

The voting machine vs. weighing machine lens

• Short-term: markets are popularity contests — fast talkers, prestigious credentials, and social momentum drive valuations
• This produces real errors: Clinkle, "professor coins," and some current AI darlings
• Long-term: every company's value is discounted future cash flows; customers must stay, pay, and renew
• The only durable moat is retention — the first renewal matters less than surviving every subsequent one
• Companies quietly compounding on small raises (and staying private) are underrepresented in the narrative precisely because they don't need to raise or go public

Why students are more sceptical than practitioners

• Harvard and MIT students encountered directly experienced the 2020–22 crypto cycle — many felt burned personally or through peers
• Silicon Valley insiders see near-universal consensus that AI is a historic moment; Cambridge students were mostly building non-AI startups
• The disconnect is real and worth noting: proximity to working AI products changes the prior dramatically
• The right question for early-career founders is not "is this a hype cycle?" but "where in the stack can I build with the least capital and the most leverage?"