What Dot.Com Bandwidth Taught Me About the AI Token Cost Panic
What Dot.Com Bandwidth Taught Me About the AI Token Cost Panic
John Honovich
Jun 5, 2026
The economic challenges I navigated at the start of my career offer a useful lens for where AI infrastructure costs are going, and how to think about them emotionally, not just technically.<br>1.5 Megabits a Second Was a Big Deal<br>A T1 line gave you 1.544 Mbps of dedicated internet access. Today, that number sounds almost comedic. On a mediocre LTE connection, your phone runs at 20 to 50 Mbps. Gigabit internet to the home, at over 600 times the capacity of a T1, costs about $60 a month.<br>But in the late 1990s, at the start of my career, a T1 was serious infrastructure. It ran about $1,000 a month. A T3, at 45 Mbps, cost tens of thousands. These were significant budget line items, and they shaped many product decisions in ways that are hard to remember now that the constraint has evaporated (or even fathom if you had not lived through it).<br>If you were building anything with real traffic, bandwidth was the wall you hit. What codec? How much compression? How much will this feature cost per user in transmission? How do we get more capacity, and how fast? There were entire engineering disciplines organized around working within constraints. The optimization work was real, the anxiety was real, and it was completely rational given the economics at the time.<br>The Constraint Went Away<br>Over the following decade, bandwidth prices cratered. The combination of massive infrastructure buildout, competition among providers, and continuing improvements in transmission technology pushed per-megabit costs down relentlessly. What had been a first-order concern for every product team became, effectively, a rounding error.<br>Few today architect a product around bandwidth scarcity. The cost is still there, technically. But it is so low relative to everything else in the stack that it no longer generally counts as a real decision variable. You do not hold a meeting about your data transmission bill the way teams in 1999 held meetings about their T3.<br>Tokens Are at the Same Early Stage<br>Anthropic's revenue has grown more than 10x in roughly the past year. That number reflects how fast token consumption is scaling across the industry, and it is the mirror image of what enterprise and developer customers are spending. For anyone building seriously on these models, the costs are real, significant, and growing.<br>The right response right now is the same one from the bandwidth era: optimize. Caching strategies, model selection, prompt efficiency, context window management, the NVIDIA and Microsoft partnership pushing custom silicon harder than ever, all of it is worth serious engineering attention. The gap between thoughtful architecture and naive architecture, in terms of token costs today, is genuinely large. The savings matter.<br>The Emotional Lesson<br>But the thing I actually took from living through the bandwidth era is not technical. It is more emotional.<br>There was real anxiety in those years about whether the economics of internet products would ever work at scale. The costs felt structural, not cyclical. Smart people built entire strategic frameworks around the assumption that bandwidth scarcity was a permanent feature of the landscape rather than a temporary condition.<br>What I learned is that you can trust the market to work this out. Competition among model providers is already pushing prices down. Custom silicon efforts from NVIDIA, Microsoft, and others are accelerating. Model efficiency is improving with every generation. The direction is not in doubt. Whether it takes three years or ten is unclear, but the trajectory is set.<br>At some point, spending billions of tokens will feel like spending billions of bits in data transmission does today. A cost, but a background one, and not something you design around.<br>What to Do With That<br>Optimize now. The savings are real, and the work has value. Just build with the confidence that the constraint is temporary, not permanent. The teams that internalized bandwidth scarcity as a first principle, rather than a current condition, built things that didn't survive the transition.<br>The ones who optimized within it while architecting for a world beyond it built things that did.
Making AI Agents Human Is Our Goal at Axamy ›
Management is broken. Let's fix it.<br>Try Axamy Free