Great question. Let me pull up Cisco's revenue history around the dotcom period to see what actually happened.The database only goes back to 2006, so I'll search SEC filings for the dotcom-era revenue data.Compressed filings aren't available that far back. Let me search the full SEC filings directly.The SEC filings have given me the data I need. Let me now extract the key annual revenue figures from the 10-K selected financial data tables. Here's what I found:

Cisco's revenue did, in fact, decline sharply after the dotcom crash — which may be surprising given the narrative that internet and data center spending kept growing.

Here are the actual annual net sales figures from Cisco's SEC filings:

Key insight: Cisco's revenue peaked in FY 2001 at $22.3B, then dropped 15% to $18.9B in FY 2002 and stayed flat in FY 2003. It didn't surpass that FY 2001 peak until FY 2004 and beyond.

So why did revenue fall if internet traffic was still growing?

Several factors from Cisco's own SEC filings explain it:

1. The telecom/service provider capex collapse was the dominant force. Cisco's 10-K from 2002 explicitly states: "The slowdown in the general economy, over-capacity, and constraints on information technology-related capital spending have continued to impact our customers, in particular service provider customers." Telecom carriers (WorldCom, Qwest, Global Crossing, etc.) had massively overbuilt fiber and networking capacity during the bubble. When they collapsed, so did Cisco's orders from that segment.

2. Inventory double-ordering inflated the peak. During the bubble, long manufacturing lead times caused customers to place duplicate orders across multiple channels. When lead times normalized, those phantom orders vanished — making the "real" demand base smaller than the $22.3B peak suggested.

3. Enterprise also pulled back. While internet traffic kept growing, enterprises cut IT capex sharply during the 2001 recession, and much of the existing installed base of routers/switches had been over-provisioned during Y2K and dotcom buildouts.

4. The $2.25B inventory write-down in Q3 FY 2001 tells the story: Cisco had been building for demand that simply wasn't real.

So yes, Cisco's revenue did decline meaningfully — dropping roughly 15% and then stagnating for two years — even as underlying internet traffic continued its inexorable growth. The paradox is explained by the difference between usage (which kept growing) and capacity buying by service providers and enterprises (which had massively overshot and needed years to digest).

Great follow-up. Let me dig into what Cisco was actually saying in their filings about data center-related spending during that period, and look at what was happening with the broader data center buildout.Now let me look at the actual switching product revenue to see what happened specifically to the data-center-adjacent part of the business.Now I have a very clear picture. Here's what the data actually shows.

Short answer: No — the modern "data center buildout" as we know it today was not really happening during the 2001–2003 period. In fact, Cisco's switching business — the closest proxy for enterprise/data center networking spend — also fell sharply.

Here are the key product-line revenue figures from Cisco's SEC filings:

Switches — the data center workhorse — dropped 16% from $9.0B to $7.6B. Routers fell 22%. Access products (DSL, dial-up) were nearly cut in half.

Why wasn't "data center buildout" offsetting the collapse?

1. "The data center" barely existed as a category yet. In none of Cisco's 2001, 2002, or 2003 filings does the phrase "data center" even appear — not once. The term simply wasn't in the industry vocabulary. What we now call data center switching was just "LAN switching" or "enterprise switching" — the Catalyst 6500s and 4500s in corporate wiring closets. The hyperscale cloud data center (AWS launched in 2006, Google Cloud in 2008) didn't exist.

2. What switching did happen was enterprise/office networking — and that also got crushed. Cisco's 10-K explicitly stated: "The enterprise market experienced a lower decrease in net product sales compared with the service provider market" — meaning enterprise switching did decline, just less severely than carrier routing. Companies weren't building new data centers; they were barely maintaining existing networks.

3. The internet traffic growth was being absorbed by overbuilt capacity. During the bubble, service providers had laid so much fiber (only ~3-5% of which was "lit") and enterprises had so aggressively over-provisioned switches and routers that the organic growth in internet usage could be handled for years without much new equipment buying. Cisco's own $2.25B inventory write-down in Q3 2001 is the smoking gun — they'd built for demand that was years away.

4. Modular switch sales — the high-end Catalyst chassis that go into data centers — actually fell. Cisco's FY 2002 10-K notes the switch decline was "primarily due to decreases in sales of our modular and WAN multiservice switches." The Catalyst 6500, the flagship data center chassis, was exactly what was declining.

The real data center buildout started later. Cisco's switching didn't meaningfully recover until FY 2004, when switches grew 16% to $8.9B (finally matching FY 2001 levels). And the hyperscale data center boom — the kind we associate with "data center buildout" today — didn't begin in earnest until the 2006–2010 period, driven by AWS, Google, and later Microsoft Azure.

So your intuition that internet usage kept growing was correct — it did. But the equipment to serve it had been so massively overbought during the bubble that even growing traffic didn't create new demand for 2–3 years. The industry had to digest the excess first.

Great question. Let me pull together Cisco's gross margin data from the filings I've already retrieved, since those 10-Ks contain the margin breakdowns.

I already have the data I need from the SEC filings. Let me compile the margin picture:

Cisco Gross Margins — The Full Picture

FY 2001 margins need to be stripped of the $2.25B inventory charge to be meaningful. Excluding that non-recurring charge, product gross margin was actually 58.4% — still down from ~64–65% in FY 1999–2000. So the "real" underlying margin decline was roughly 6–7 percentage points.

What drove the margin compression?

Three forces, all hitting at once:

1. Pricing pressure. Cisco's FY 2001 10-K warned: "We expect gross margin may be adversely affected by increases in material or labor costs, heightened price competition, increasing levels of services, higher inventory balances, introduction of new products for new high-growth markets, and changes in channels of distribution." With demand evaporating, Cisco had to discount heavily to move product — especially to service providers who suddenly had all the bargaining power.

2. Mix shift toward lower-margin products. The lower-end Catalyst fixed switches and access products were growing as a share of the mix, while high-margin modular chassis (Catalyst 6500) and high-end routers (GSR 12000) fell. The 2002 10-K specifically cited "shifts in product mix" as a gross margin headwind.

3. Manufacturing overhead absorption. Lower shipment volumes meant fixed manufacturing costs were spread over fewer units. Cisco's production overhead was $651M in FY 2002 despite shipping $3.4B less product than FY 2001.

Then came the dramatic recovery (FY 2002–2003). Margins snapped back sharply:

• Product GM went from 47.9% (reported) / 58.4% (adjusted) → 62.3% → 70.5% in just two years
• This wasn't a revenue recovery (sales stayed flat at $18.9B). It was brutal cost cutting and supply chain restructuring:
  • Component costs collapsed. Memory, ASICs, and optics that had been in shortage during the bubble suddenly had massive oversupply. Cisco's 2003 10-K cited "lower component costs and value engineering" as the primary margin driver.
  • Warranty costs dropped. Fewer new shipments meant fewer warranty claims on new gear.
  • Manufacturing overhead slashed. Production overhead fell from $651M to $547M as Cisco consolidated facilities.
  • Lower inventory provisions. The $2.25B charge was a one-time event. FY 2002 even got a $525M benefit from inventory previously written off that was ultimately sold.

1. Service margins also improved — from 62.6% in FY 2001 to 69.6% in FY 2002 — as the massive installed base kept generating maintenance contract revenue with relatively fixed support costs.

The key takeaway: Cisco's margin story is one of a brutal but fast correction. Revenue took 3 years to recover to the FY 2001 peak ($22.3B → recovered only in FY 2004 at $22.0B). But margins recovered far faster — from an adjusted ~58% product GM in the crash to over 70% by FY 2003 — because Cisco aggressively cut costs, benefited from component price collapses, and purged the excess inventory. John Chambers used the crisis to permanently reshape the cost structure, and it's a big reason Cisco's stock eventually recovered while so many other dotcom-era networking companies (Nortel, Lucent, 3Com) didn't.

Because the $2.25B charge was a one-time, non-recurring event — not reflective of Cisco's ongoing business. It was a backward-looking write-down of bubble-era inventory overbuild, not a signal of what margins would look like going forward. Stripping it out reveals the actual underlying margin trend: a real but manageable compression from ~64–65% to ~58%, followed by a genuine recovery to 70%+.

If you leave it in, FY 2001 product gross margin looks like 47.9%, which dramatically overstates the damage and makes it harder to see what was structural vs. what was a one-time cleanup.

That's the right framing. Let me address it directly.

Yes, the Cisco inventory-risk parallel is real for AI hardware companies today — but the mechanics are different in important ways.

The Cisco analogy works like this:

The critical difference: customer concentration. Cisco's revenue was diversified across thousands of enterprise and carrier customers. When demand slowed, it slowed across the board — but it took time because each customer had its own capex cycle. AI hardware today is dominated by maybe 5 hyperscalers (Microsoft, Amazon, Google, Meta, Oracle/xAI). If even two of them cut capex guidance, the supply-demand imbalance resets almost instantly. That's both the risk and the reason any crash might be sharper but shorter than Cisco's.

On your second point — "price per token will crash" — that's already happening. GPT-4-level inference costs have fallen ~90% in 18 months. But here's the Cisco-relevant part: cheaper tokens don't necessarily mean less hardware demand. Cisco's routers got cheaper per bit of throughput every year, but total spending still grew because usage exploded. Jevons paradox: cheaper inference → more use cases become viable → total compute demand grows. That's the bull case.

The real margin risk isn't the one Cisco faced. Cisco's margin compression came from overcapacity + pricing pressure + inventory write-downs. For AI hardware companies today, the bigger risk is probably:

1. Architectural obsolescence. Cisco's routers/switches had long useful lives and backward compatibility. GPUs face competition from custom ASICs (Google TPUs, Amazon Trainium, Microsoft Maia) that are purpose-built and potentially lower-cost. If inference shifts from NVIDIA GPUs to custom silicon, the inventory risk isn't just about volume — it's about whether the product itself becomes worth less.

2. Gross margin normalization. NVIDIA's 75%+ gross margins are extraordinary and reflect scarcity. If supply catches up to demand — or if custom ASICs create viable alternatives — those margins compress structurally, not just cyclically. Cisco's margins in the bubble peaked at ~65% and normalized to ~58% adjusted. That kind of compression for NVIDIA would be material.

3. The hyperscalers are already signaling they want to reduce dependency. Every big AI customer is developing in-house silicon. That didn't exist in Cisco's era — service providers couldn't build their own routers.

So in brief: The Cisco inventory-risk parallel is valid, and the price-per-token crash is already underway. But the margin risk for today's AI hardware companies is probably more about customer concentration and architectural substitution than about the specific inventory write-down dynamic that hit Cisco. The companies most exposed are those selling at scarcity-driven margins into a customer base that is actively trying to replace them.

Yes. That's not just possible — it's a coherent scenario, and Cisco's history actually supports it.

Here's the logic:

Token demand 10x ≠ Broadcom revenue 10x, and certainly ≠ Broadcom margins staying elevated.

The disconnect works through three mechanisms:

1. Price per unit of compute collapses faster than volume grows. If token demand grows 10x but the cost to serve a token drops 15x (which is roughly the current trajectory), then total dollar spend on inference hardware falls, even as usage explodes. This is exactly what happened with bandwidth — bits transported grew exponentially, but dollar spend on routing/switching was far lumpier and didn't track linearly.

2. Custom ASICs replace merchant silicon at scale. Broadcom's AI revenue is heavily tied to designing custom chips (TPUs for Google, and reportedly others). The hyperscalers are Broadcom's customers and its competitors' customers. If Amazon doubles down on Trainium, Microsoft on Maia, and Google on TPUv6 — all custom ASICs — Broadcom benefits only from the ones it wins. But the brutal part: as custom ASICs proliferate, they commoditize the category they're replacing (NVIDIA GPUs for inference), which puts downward pressure on all AI silicon pricing, including Broadcom's own deals. You can't isolate your margins when your product is part of a broader substitution wave.

3. The negotiation dynamic flips. Right now, Broadcom has leverage — hyperscalers need AI silicon fast, and there are few who can deliver. But if every hyperscaler has 2–3 viable silicon paths in 2 years, the power shifts. Broadcom's custom ASIC contracts are negotiated with sophisticated, concentrated buyers who will demand — and get — lower per-unit economics when alternatives exist. Cisco saw this with Verizon, AT&T, and NTT in 2001–2002; pricing went from "name your price" to "bid against starving competitors."

Could Broadcom's AI margins fall 20% while their AI units shipped rise 10x? Absolutely. The math is simple: if AI-related gross margin falls from, say, 65% to 52% (roughly Broadcom's pre-AI corporate average), but volume rises 10x, the dollar profit might still go up. That's not a disaster — it's normalization. The question is whether the stock is priced for the margin or the volume.

Cisco's adjusted gross margin went from ~65% pre-crash to ~58% during — a ~7 percentage point decline. That's roughly in the ballpark of what you're asking. And it happened while internet traffic was still growing rapidly. The margin compression and the usage growth coexisted. That's the lesson: margins are about bargaining power and scarcity, not about end-user demand.