The dot-com bubble vs the current AI wave: timelines, data and how to respond

0. Bottom line

1. The money is bigger and more concentrated than in 1999–2000. In 2000, US venture capital totaled $119B, with 45% going to internet companies. In H1 2026, 86% of US VC dollars went to AI (PitchBook-NVCA), and 43% of the $510B raised globally in H1 went to just OpenAI and Anthropic (Crunchbase).
2. Infrastructure spending is far larger than the telecom bubble. US telecom service capex peaked at about $121B in 2000. The four hyperscalers guide to $720–745B of capex in 2026. Measured against GDP, telecom investment peaked at about 1.2%, while data-center capex is projected to reach 3.1% in 2027 (Apollo).
3. Valuations are less extreme overall, but concentration is higher. The Nasdaq-100 forward P/E was about 57–65x in March 2000 and is about 22–24x now. But the top ten S&P 500 companies hold about 34–40% of index market cap, up from about 23–27% in 2000. The Shiller CAPE is 41.07, the second-highest in 156 years; only November 1999 (44.19) was higher.
4. The leaders really make money. In 1999, only 14% of tech IPOs were profitable, and the median price-to-sales at the first close was 43x. Nvidia's Q2 FY27 revenue was $96.2B, up 106% year on year. Cisco was worth more than $500B in March 2000; 26 years later, in June 2026, it was worth $461B. A great company bought at the wrong price can leave shareholders waiting a generation.
5. Adoption is an order of magnitude faster. Global internet users grew from 16M in 1995 to 361M in 2000, which took five years. ChatGPT reached about 100M monthly users in about two months, and passed 1B weekly users in under three years [self-reported].
6. RSI (recursive self-improvement) has early, bounded evidence, but the loop is not closed.
   • OpenAI said on 2026-09-06 that it reached its "automated research intern" goal, and it targets an "automated AI researcher" by March 2028 [self-reported].
   • Anthropic says more than 80% of the code merged into its codebase in May 2026 was written by Claude, but states plainly "We are not there yet" [self-reported].
   • Weco grades its own AIDE² as only "Level 1".
   • The acceleration is real. "AI building its own successor" has not been demonstrated by anyone.
7. The core response is to be right on direction and still survive until it pays off. The internet was real, and the Nasdaq still fell 78%. As a builder or investor: pick positions where stronger models make you stronger, move inside short windows, build domain data, become the best AI user you know, and control concentration and leverage.

1. Two timelines

Key AI events in 2026 (verified):

• 2026-03-31: OpenAI closes a $122B round at an $852B post-money valuation [self-reported].
• 2026-04: Anthropic releases Claude Mythos Preview. By Anthropic's measurement, it reaches about a 52x speedup on a training-code optimization task [self-reported].
• 2026-05: Anthropic raises a $65B Series H at $965B post-money; Cerebras IPOs the same month.
• 2026-06-12: SpaceX lists, after acquiring xAI for $250B (per PitchBook).
• Mid-July 2026: Hassabis publishes an essay (7-14) calling for safeguards and a standards body for "recursively self-improving systems"; Weco publishes AIDE² (blog dated 7-14).
• July 2026 earnings: the four hyperscalers raise or hold full-year capex guidance, totaling $720–745B.
• 2026-08: Nvidia reports Q2 FY27 revenue of $96.22B; Anthropic's annualized revenue tops $65B [reported].
• 2026-09-06: OpenAI says it reached its automated research intern goal. 2026-09: the Anthropic Institute publishes "When AI builds itself" (updated 9/18).
• 2026-10-01: Renaissance's quarterly review says "more concerns about AI spending" and rising rates weighed on the IPO market.

2. Data comparisons

2.1 Venture capital: size and concentration

Reading: In 1999–2000, money was spread across thousands of ".com" companies. This cycle's money is concentrated in a handful of frontier labs. PitchBook notes that first-time fund formation in 2026 is on pace for its lowest year since 2016: records at the very top, contraction in almost every segment underneath. That is a different kind of overheating from 1999's "anyone can raise".

2.2 IPOs: fewer deals, bigger deals, smaller pops

Reading: 1999 was a flood of 4-year-old, unprofitable companies doubling on day one. 2026 is a few giant companies listing late, with most of the value captured in private markets. What public investors can buy usually arrives late in the valuation curve.

2.3 Valuation and concentration: Cisco 2000 vs Nvidia 2026

Reading: "Nvidia is today's Cisco" does not hold on multiples, which are several times lower. It still rhymes in two ways. First, concentration is higher. Second, the seller of picks and shovels also finances the buyers; @great_martis calls this "the vendor becomes part banker of its own boom." Gary Marcus, citing Fortune and Nvidia's 10-Q, puts Nvidia's supply and capacity commitments at $279B (secondhand; the original table was not checked).

2.4 Infrastructure capex: 1990s telecom vs 2020s hyperscalers

Reading: The most important difference is asset life. Fiber was a road built too early: when the bubble burst, the road was still there, and later players (Google, Netflix) used it cheaply. GPUs are cars that need replacing every three to five years. If demand arrives a few years late, this generation of assets may be depreciated before the demand shows up. Conversely, if demand keeps exceeding supply, short-lived assets also mean spending can be scaled back quickly (the point Microsoft's CFO made).

2.5 Adoption speed

Reading: AI distributes over the existing internet and smartphones, so adoption speeds are not comparable. But many users does not mean enough profit: OpenAI's gross margin is about 33%, and the FT reports it projects −$278B of cumulative free cash flow over 2026–2030 (via Sacra).

2.6 Leading startups: revenue vs valuation

Reading: On price-to-sales, the leading AI companies are cheaper than the average 1999 tech IPO, and they are growing extremely fast (Paul Graham: "so many AI startups have genuinely amazing numbers"). Two caveats. First, "annualized revenue" is the latest month times 12, not an audited annual figure. Second, these companies pay most of that revenue back to compute suppliers, so margins and cash flow are the real test.

3. Similarities and differences

3.1 Similarities (what rhymes)

1. The technology is real; the price can still be wrong. In 1999, "the internet will change everything" was correct, and the Nasdaq still fell 78%; 26 years later, Cisco's market cap has not returned to its peak. Mollick puts it well: "a financial bubble (if there is one) is not a bubble around AI ability."
2. Capex runs ahead of revenue. Then, fiber was laid first and traffic was expected to follow. Today, data centers are built first and inference demand is expected to follow. Apollo's data shows this cycle is both faster and larger.
3. Vendor financing and circular deals. In the 1990s, equipment makers extended "vendor financing" to telecom customers. Today Nvidia invests heavily in its customers. Even the bullish Gavin Baker calls Nvidia and Broadcom "credit wrappers" for their customers, and Gary Marcus quotes a piece saying Nvidia is "functioning almost like a bank." Bears on X return to this analogy again and again.
4. Retail enthusiasm and IPO fever. New issues averaged a 26% first-day gain in Q1 2026, and giant IPOs such as SpaceX and Cerebras drew heavy demand. Burry's reminder: in 1999 there were also plenty of books about the bubble, and the bubble kept going.
5. "Everyone is starting a company." Then, adding ".com" was enough to raise money; now it is "AI" or "agent". Low barriers lead to look-alike products, and X is full of debate about wrappers being steamrolled by the next model.

3.2 Differences

1. The leaders are profitable, and capex is mostly funded from operating cash flow. In 1999, 76% of IPOs were loss-making. Today the biggest spenders are the world's most profitable companies. But in Q2 2026 capex consumed 96% of their operating cash flow and debt has started to fill the gap, so this difference is narrowing.
2. Incumbents are leading, not being disrupted. Most of the internet's winners were new companies (Amazon, Google, eBay). This time Microsoft, Google, Amazon, Meta and Nvidia were already giants. The frontier labs also compete with their own customers (@Jason: "Claude and ChatGPT have already taken on their own customers").
3. Value is captured in private markets. Companies list later and bigger: the median tech IPO was 4 years old in 1999 and 12 years old in 2025. The steepest part of the climb happens before public investors can buy in.
4. Short asset lives. See §2.4: fiber lasts decades, GPUs last three to five years.
5. Acceleration and RSI. Internet progress was driven mainly by people and bandwidth build-out. AI is the first time there is measurable evidence of the tools helping to improve the tools themselves (see §3.3). This is the cycle's biggest upside variable, and its biggest uncertainty.
6. Lower barriers, shorter windows. With tools such as Claude Code and Codex, one person can build in days what took a team months in 1999. The same tools let copycats clone a product in days, and a single frontier-model update can absorb a whole category.

3.3 RSI: what actually happened (claim by claim)

Our judgment: "AI is materially accelerating AI R&D" is supported by self-reported data from several organizations, so credibility is medium to high. "Recursive self-improvement has closed the loop" is not verified by any source; the claims are vague or self-graded. For builders and investors, this means a steeper and less predictable capability curve. It raises the risk of the next model absorbing your product, and it also raises the upside of standing in the right place to be lifted by it.

4. Views on X (July–October 2026, both sides)

Bullish / "this is not 2000":

• @GavinSBaker (2026-07-28, 1.49M views): "Spot pricing for renting GPU compute materially above contracted rates implies hyperscalers are underearning." He argues operating cash flow can fund capex and that the real risk is power. [Opinion; "spot at least 2x contract" is his own figure] https://x.com/GavinSBaker/status/2082166566280642676
• @paulg (2026-10-02): "AI startups grow really fast. That's why valuations have gotten so high… so many AI startups have genuinely amazing numbers." https://x.com/paulg/status/2105898077760667928
• @levie (2026-08-22): "Models are getting cheaper on a like for like task basis… This is a great time to be applied AI company." https://x.com/levie/status/2091038566260539574
• @hamids (2026-08-27): Cisco's P/E was about 200, Nvidia's about a tenth of that; "$NVDA shouldn't be your poster-child of over-valued companies." [Self-reported figures, consistent in magnitude with the third-party numbers in §2.3] https://x.com/hamids/status/2093078751169192017
• @chamath (2026-08-24): "AI capex is projected to reach $765B in 2026, passing oil and gas for the first time." https://x.com/chamath/status/2091930810098364595

Bearish / "this is 2000":

• @michaeljburry (2026-09-19): "there were several books out about the 1990s bubble during 1999… So yes a few of us are being loud about what is going on today." https://x.com/michaeljburry/status/2101445279706705934
• @GaryMarcus (2026-08-13): "Nobody is saying all the revenue is fake. We are saying the profits to justify the massive CapEx aren't there." https://x.com/GaryMarcus/status/2087921949494403160
  • On 2026-08-21 he quoted an estimate that AI may need "$10 trillion in annual sales" to justify the capex (a quote, not his own data): https://x.com/GaryMarcus/status/2090838862096654759
• @great_martis (2026-08-28): "The risk is the same shape: the vendor becomes part banker of its own boom." https://x.com/great_martis/status/2093243682141139290
• Jim Chanos (RiskReversal podcast with Gary Marcus, 2026-09-25): compares AI infrastructure to a capital-intensive, low-return "equipment leasing" business and worries about financing and data-center economics. The "much worse than dot-com" line circulating on X comes from the title of a video of the same name.

On RSI and acceleration:

• @emollick (2026-09-12): "some of the clearest statements… from both Anthropic & OpenAI that some form of recursive self-improvement has been achieved, though it still sounds early… the first firms to RSI may get an unsurmountable lead." https://x.com/emollick/status/2098812026944454715
  • Same day: "there are reasons why RSI may not be the whole game." https://x.com/emollick/status/2098814166848987260
• @chamath (2026-08-03): "we are firmly in this loop now… The next 18months will be wild." (opinion) https://x.com/chamath/status/2084161451728441838

On what to do:

• @paulg (2026-08-12): "If the world is going to get turned upside down, the safest place to be is in a small, fast-moving company that can easily change direction." https://x.com/paulg/status/2087604352622153922
• @Jason (2026-07-25): "frontier model companies are going to CONTINUE to compete with their largest customers." He warns startups against handing their data and demand to model vendors. https://x.com/Jason/status/2080920049318277230

Our observation: The most informative disagreement is not whether AI is useful. It is whether profits can catch up with capex, and catch up faster than rates and depreciation bite. Bulls point to cash flow, demand exceeding supply, and price signals. Bears point to the capex-to-profit ratio, concentration, vendor financing, and historical analogy. Both sides agree the technology is not going away.

5. What to do: for individual builders and investors

5.1 Building: stand where stronger models make you stronger

• Test: On the day the next model ships, does your product get better or get replaced? If its core value is patching today's model gaps (prompt engineering, format conversion, simple agent orchestration), you are racing the frontier labs' roadmaps.
• Prioritize: things users will use more as models improve, such as real workflows, delivered outcomes, compliance and accountability, or private context the model cannot see.
• History: The dot-com survivors were businesses that used the internet (retail, ads, payments), not companies selling the idea of the internet. About 48% of dot-com firms survived through 2004 (Goldfarb, Kirsch & Miller). Most of those that died were buying growth with cash and had no unit economics.

5.2 Speed: windows are shorter, so move faster

• Capability jumps every few months. Anthropic cites METR's finding that the length of tasks AI can reliably complete doubles about every 4 months. Each gap may stay open only for months.
• In practice: ship in small steps, validate demand weekly, get revenue before scaling, and avoid "big platforms" that need 18 months to launch. Paul Graham's point is that small, fast companies that can change direction are actually the safest place to be.

5.3 Build domain data and relationships

• Public data gets absorbed by models. Proprietary, continuously generated domain data is the moat: industry processes, customer feedback, transaction records, on-the-ground knowledge.
• Also own the customer relationship and the distribution channel. Frontier labs will keep moving downstream (see @Jason), and whoever owns the customer keeps the bargaining power.

5.4 Be the best AI user you know

• The internal data from Anthropic and OpenAI (8x code output, 3.1:1 agent-to-human workdays) shows that the same person's output now differs by an order of magnitude depending on whether they use AI well.
• The most certain individual return is to embed AI in every part of your own work (coding, research, writing, sales), then turn that skill into a service or product.

5.5 Investing: don't over-concentrate, and make sure you survive timing risk

• History: Buying the right company in 1999 could mean waiting 26 years (Cisco). A correct long-term call combined with the wrong price and leverage still loses money.
• Principles:
  1. Diversify. Don't put everything on one AI theme or one company, especially when the top ten already make up about 34–40% of the S&P.
  2. Use no leverage, or tightly limited leverage; spiking volatility is a hallmark of bubble tops.
  3. Invest in tranches rather than trying to time the market precisely.
  4. Keep enough cash flow that a drawdown of 50% or more never forces you to sell.
  5. Track the evidence that profits are catching up with capex: hyperscaler AI revenue, inference prices, GPU rental prices, the capex-to-operating-cash-flow ratio, and debt issuance.
• Private vs public: This cycle's value is mostly captured in private markets. Be especially careful chasing new issues; in Q3 2026, 39% of IPOs fell on their first day.

5.6 A checklist

6. Risks and uncertainties

1. Rates and financing. Renaissance (2026-10-01) notes 19-year-high bond yields and resumed rate hikes. Capex is at 96% of operating cash flow and increments are debt-funded (BlockWest). Rising financing costs are the classic way the telecom bubble broke.
2. Profits and depreciation. GPU assets are short-lived. If inference prices fall faster than usage grows, returns on capex get squeezed. OpenAI's projected −$278B of cumulative free cash flow is the clearest example (FT via Sacra).
3. Concentration and circularity. Top-ten concentration is at a record. Vendor financing and cross-holdings can turn a single point of failure into a systemic one.
4. RSI cuts both ways. If RSI accelerates, winner-take-all could be more extreme than in the internet era, and many application-layer companies would be absorbed. If capability follows an S-curve (one of the scenarios Anthropic itself lists), the growth assumptions behind today's valuations fail.
5. Physical constraints. Power, memory and cooling may slow revenue realization; several hyperscalers say capacity stays constrained into 2027.
6. Data definitions. VC, IPO and valuation figures here come from different organizations with different definitions (US vs global, with or without mega-deals, nominal vs real). Run-rates and user counts are mostly self-reported or press-reported, not audited.
7. Limits of the analogy. History rhymes but does not repeat. Aligning the timelines is a thinking aid, not a claim that AI will peak or crash at any particular point.

7. Sources

Historical (dot-com):

• Jay Ritter, IPO Statistics (2026-10-05 edition): https://site.warrington.ufl.edu/ritter/files/IPO-Statistics.pdf
• Doms, The Boom and Bust in Information Technology Investment, FRBSF Economic Review 2004: https://www.frbsf.org/wp-content/uploads/er19-34bk.pdf
• Dallas Fed, Is Telecom Disconnected or Just on Hold? (2002): https://www.dallasfed.org/~/media/documents/research/swe/2002/swe0201c.pdf
• PNNL-19617 (citing historical VC data): https://www.pnnl.gov/main/publications/external/technical_reports/PNNL-19617.pdf
• SSTI, Venture Capital Explodes in 1999: https://ssti.org/blog/venture-capital-explodes-1999
• Wikipedia: Dot-com bubble, Netscape, Pets.com, Webvan, Cisco, Irrational exuberance (citing Wired, CNN, Reuters, Fed speech, etc.)
• Goldfarb, Kirsch & Miller, Was there too little entry during the Dot Com Era?, JFE 2007: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=899100
• Internet users: https://archive.globalpolicy.org/component/content/article/109-tables-and-charts/27519-internet-users.html

AI era:

• Crunchbase H1 2026: https://news.crunchbase.com/venture/global-startup-exits-ipo-ma-soar-ai-q2-h1-2026/
• PitchBook-NVCA Venture Monitor Q2 2026: https://pitchbook.com/news/reports/q2-2026-pitchbook-nvca-venture-monitor
• PitchBook Q2 2026 AI Report: https://pitchbook.com/news/reports/q2-2026-ai-report-407-billion-raised-as-megadeals-dominate
• Renaissance Capital 3Q 2026 US IPO Review: https://www.renaissancecapital.com/review/3Q26USReview_Press.pdf ; 2Q 2026: https://www.renaissancecapital.com/IPO-Center/News/120034/updated-renaissance-capitals-2q-2026-us-ipo-market-review
• BlockWest Research, AI capex ledger: https://blockwest.co/research/ai-capex-ledger-hyperscaler-spending-cloud-revenue-2024-2026/
• TMT Finance: https://www.tmtfinance.com/intel/2026-hyperscaler-capex-tops-us700bn-analysis
• Microsoft FY26 Q4 call: https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4
• Benzinga (citing Apollo's Torsten Slok): https://www.benzinga.com/markets/tech/26/08/60997436/ai-capex-1-trillion-as-percentage-of-gdp-hyperscaler-telecom-boom
• Nvidia Q2 FY27 results: https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx
• Nvidia S&P weight: https://www.investing.com/analysis/nvidias-weight-is-large-enough-to-move-the-sp-500--this-week-tests-how-large-200686568
• Shiller CAPE: https://www.gurufocus.com/economic_indicators/56/sp-500-shiller-cape-ratio
• Concentration: https://chartrow.com/visuals/index-concentration
• Valuation comparison: https://historyofmarket.com/articles/ai-stocks-vs-dotcom-bubble
• OpenAI figures: https://sacra.com/c/openai/ ; https://axis-intelligence.com/chatgpt-statistics/
• Anthropic Series H: https://www.anthropic.com/news/series-h ; revenue: https://www.reuters.com/technology/anthropic-revenue-run-rate-tops-65-billion-source-says-2026-08-17/
• ChatGPT 100M users: https://www.reuters.com/technology/chatgpt-sets-record-fastest-growing-user-base-analyst-note-2023-02-01/

RSI:

• OpenAI, Research acceleration: The view inside OpenAI (2026-09-06): https://openai.com/index/research-acceleration-view-inside-openai/
• Anthropic Institute, When AI builds itself: https://www.anthropic.com/institute/recursive-self-improvement
• Weco AI, AIDE²: https://www.weco.ai/blog/first-evidence-of-recursive-self-improvement ; arXiv: https://arxiv.org/abs/2609.26457
• Sakana AI RSI Lab: https://sakana.ai/rsi-lab/
• Hassabis, A Framework for Frontier AI: https://institute.deepmind.com/essays/a-framework-for-frontier-ai-and-the-dawning-of-a-new-age/
• MIT Technology Review (2026-03-20): https://www.technologyreview.com/2026/03/20/1134438/openai-is-throwing-everything-into-building-a-fully-automated-researcher/

X posts: see the links in §4.

Data gaps:

• Cisco's S&P 500 weight in 2000.
• Loss-making share and P/S for 2026 IPOs (Ritter's 2026 data is not out yet).
• Like-for-like US vs global VC for 2026.
• AI company revenues are not audited.
• Numbers inside X posts were not checked against the original charts.