Five tests from 2000

What separated the companies that disappeared in 2000–02 from those that made it through, turned into tests on the latest filings. A company that trips a test gets a flag. Fewer flags means more staying power in a bust; it says nothing about whether the share price is cheap or will rise.

Physical infrastructure (data centres, power)

In 2000Fibre networks and carriers

The layer in fibre's seat. But power, land and grid connections last decades and cannot be added quickly. Even if overbuilt, they are the kind of asset latecomers end up using. The risk sits with debt-funded builders and with the creditworthiness of the AI tenants on the other side of long contracts.

CompanyBusinessRevenue (4Q)GrowthOp. marginCapex ÷ OCFNet cashP/EP/Svs 52w highFlags
VertivVRTPower and cooling for data centres$11bn+26%19%13%-$0.1bn56x8.5x-33%None
VistraVSTTexas-centred generator betting on data centre demand$19bn+4%19%56%-$19bn21x2.4x-34%Heavy debt
Constellation EnergyCEGNuclear-heavy generator with long-term data centre contracts$31bn+26%15%93%-$19bn27x3.0x-35%Heavy debt

Chips and networking gear

In 2000Telecom gear (Cisco, Lucent, Nortel)

Cisco's seat in 2000. It is the most profitable layer today, but its revenue is its customers' capex, and it stops when they stop. Even the pick-and-shovel survivors of 2000 lost 80–90%. Investing in customers to support demand has the same shape as Lucent's and Nortel's customer financing.

CompanyBusinessRevenue (4Q)GrowthOp. marginCapex ÷ OCFNet cashP/EP/Svs 52w highFlags
NVIDIANVDALeading AI GPU maker; in Cisco's 2000 seat$303bn+104%65%5%—28x17.9x-5%None
AMDAMDGPUs and CPUs; gave OpenAI share warrants alongside a large order$41bn+49%16%17%$9.9bn160x24.9x0%Richly priced
BroadcomAVGOCustom AI chips for hyperscalers and networking silicon$89bn+56%48%3%-$35bn44x18.9x-27%None
MicronMUHigh-bandwidth memory for AI$90bn+188%66%49%$21bn24x13.5x-11%None
Arista NetworksANETData centre switches; the closest business to Cisco's$11bn+33%43%—$13bn64x24.7x-2%Richly priced

Cloud and compute landlords

In 2000Hosting and data centres (Exodus etc.)

The main spenders. Giants with profitable core businesses can cut back and ride out a bust, as Microsoft and Oracle did in 2000. The ones that look like the 2000 carriers are the debt-funded GPU clouds and anyone spending beyond their earnings.

CompanyBusinessRevenue (4Q)GrowthOp. marginCapex ÷ OCFNet cashP/EP/Svs 52w highFlags
MicrosoftMSFTAzure; a major OpenAI shareholder$332bn+18%47%63%$37bn29x11.6x-5%None
AmazonAMZNAWS; an investor in Anthropic$776bn+16%12%107%-$10bn20x3.5x-12%Negative FCF
AlphabetGOOGLGoogle Cloud, its own chips (TPU) and models (Gemini)$446bn+20%33%71%$142bn17x9.4x-15%None
OracleORCLReported ~$300bn five-year compute deal with OpenAI (Sept 2025)$72bn+22%32%161%-$92bn22x5.8x-56%Negative FCFAsset-heavy
CoreWeaveCRWVGPU cloud upstart funding its fleet largely with debt$7.6bn+115%-3%298%-$19bnLoss6.4x-39%Negative FCFOperating lossAsset-heavy

Services and applications

In 2000Portals and e-commerce (Yahoo, Amazon, Pets.com)

In 2000 the eventual winners came from this layer, because they got to use the compute that became cheap after the bust. The winners may still be small, private or not yet founded (Google started in 1998, Facebook in 2004). The model makers (OpenAI, Anthropic) are private and not in the figures here.

CompanyBusinessRevenue (4Q)GrowthOp. marginCapex ÷ OCFNet cashP/EP/Svs 52w highFlags
Meta PlatformsMETAUses AI in ads; its own models (Llama) and giant data centres$228bn+28%38%69%$6.6bn28x8.5x-3%None
SalesforceCRMBusiness software selling AI agents on top$44bn+11%20%4%-$28bn20x4.4x-12%None
ServiceNowNOWWorkflow automation software$15bn+22%11%14%-$0.8bn84x9.5x-28%Richly priced
AdobeADBECreative software; also exposed to generative AI substitutes$26bn+15%36%2%$0.9bn13x3.5x-35%None
PalantirPLTRData analytics and AI deployment$6.2bn+79%43%1%$9.4bn151x74.0x-8%Richly priced

How to read this

Figures are the latest four quarters reported to the SEC (fiscal periods ending May 2026 to Aug 2026). Net cash is cash and short-term investments minus borrowings; lease liabilities are not counted as debt. P/E (trailing four quarters of net income) and P/S use the latest share price and share count; forward earnings estimates are not free to obtain, so P/E is backward-looking. Foreign companies (TSMC, ASML, Japanese names) and private model makers are not included.

This is not investment advice. The flags describe how a company would weather a bust, not whether its shares are a buy.

Updated 28 Sep 2026 10:50 JST · prices to 25 Sep 2026