We've spent ~30 years optimizing for throughput: faster radios, fatter pipes, lower latency. Then we got bandwidth. And we spent it. The median desktop page increased ~120% over 10 years (and mobile even more) to around 2,652 KB desktop and 2,311 KB mobile 1. And we mostly used it to ship heavier pages, more autoplay video, infinite scroll.

APIs

Today, a few kilobytes of structured text from your favorite BigLab can already contain a summary, a plan, a decision, a fix. When you include HTTP headers and a small JSON envelope:

  • Short queries (simple lookup): ~1.6 KB
  • Moderate queries (summary): ~5 KB
  • Longer queries (complex reasoning): ~12.5 KB

On average, the amount of useful 2 tokens received from frontier models has stayed roughly the same. It's what tokens are streamed that makes the difference. So when are we getting our Hassabis tokens? Our Ilya tokens? Whatever - the economic point is that intelligence compresses and the usefulness packed into each token increases by the month (or week).

Outside the bubble

At the same time, the global connectivity baseline still roughly looks like this (2022) 3:

  • 2.6 billion people remain offline
  • Globally monthly average data use was 257 GB per fixed-broadband subscription vs 11 GB per mobile-broadband subscription
  • In low-income countries monthly fixed-broadband traffic averaged 161 GB, compared to 1 GB for mobile

Now, AGI might still be A Few Years Away™. But as the cost curve keeps collapsing, more of that pipe becomes economically viable. Altman's public framing is basically "cost drops insanely fast," pointing to large per-token price declines across model generations 4, and current API pricing is explicitly per-token 5. Not to mention the progress of Chinese models, serving output tokens that are roughly 97% cheaper than their OpenAI counterparts 6.

We won't close the digital divide in time. But here's the thing: we might not need to. A farmer in rural Malawi. A fisherman in the Amazon basin with a Nokia and a text box connected to a patchy 2G signal. You can in theory access the same frontier intelligence as someone on fiber. The questions are 1) how one brings the endpoint to the billions of people without one (maybe something like 7), 2) if they had it, would they know how it could be useful to them? Would it be useful at all?

The curve

                 Value per byte

                       |                                              /
                       |                                             /
                       |                                            /
                       |     /\                                    /
                       |    /  \                                  /
                       |   /    \_________________        _______/
                       |  /                       \______/
                       | /
                       +─────────────────────────────────────────────────→ Time
                         1990s       2010s         2020s      AGI era
        
                      Dense text   Bloated web   AI slop   Dense intelligence
                      (KB matter)  (GB wasted)   (GB slop)    (KB > GB)

Looking past the decade of AI slop that's ahead of us, the value per byte on the web will increase again and certain KBs will matter more than most GBs today. And if you truly care about infrastructure and global reach, you don't ship faster GB/s. You ship better KB/s.

Footnotes

  1. HTTP Archive Web Almanac 2024 - Page Weight. ↩

  2. Not counting CoT. ↩

  3. Internet Society Pulse - One-Third of the Global Population Remains Offline. ↩

  4. Outlook Business - AI usage cost falls 10x every 12 months, says OpenAI CEO Sam Altman. ↩

  5. OpenAI API Pricing. ↩

  6. LLM API Pricing Comparison 2025 - Intuition Labs. ↩

  7. Projects like Meshtastic—open-source LoRa mesh networks that work off-grid—show that connectivity doesn't require traditional infrastructure. A few $30 devices can create a network spanning 2-5km or upwards of 100km when used via mesh (record: 331km). ↩