[Chapter-delegates] The AI revolution threatens a ‘token divide’

[Glenn McKnight]

The AI revolution threatens a ‘token divide’
Tom Wheeler
https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid

The internet revolution was accompanied by a digital divide. Now, the
artificial intelligence (AI) revolution threatens a token divide.

Access to the internet opened new possibilities for everyone. The eighth
grader doing her homework and the local business serving its customers both
benefited. Broadband connections gave them information. Now, AI offers them
intelligence.

Tokens <https://blogs.nvidia.com/blog/ai-tokens-explained/> are the unit
through which intelligence is organized, measured, and sold. Access to
tokens governs access to intelligence.

The digital divide was about access to and affordability of internet
connections. The token divide retains those challenges and adds another:
the constant expansion in what AI can do. Closing the token divide requires
access and affordability, all while keeping pace with an accelerating
frontier of capability.

Affordability of tokens determines the quantity and quality of intelligence
that can be obtained. The accelerating advance of AI capabilities deepens
the divide because affordable intelligence is only meaningful if it keeps
pace with the advancing frontier. In the end, the token divide asks whether
access to yesterday’s model is meaningful access at all.
Back to top
<https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid>
The token tollbooth

Today’s generative AI models process inputs and generate output through
computational units called tokens. The diverse applications of AI models,
ranging across text, data, and images, are all token-based. In
English-language text-based applications, a token averages roughly
three-quarters of a word. By breaking language into tokens, an AI model can
mathematically analyze relationships among pieces of language and generate
an intelligent response.

Every query (a “prompt” in AI-speak) to an AI model must be processed as
tokens, and every response must be generated as tokens. The longer the
prompt and response, and the more extensive the reasoning the model
performs, the greater the consumption of tokens and compute capacity.
Regardless of whether the user sees a per-token charge, someone bears the
cost of their processing.

Advanced reasoning models can consume substantial computation before they
produce any output. That compute must ultimately be paid for by the user,
the provider, or both. The growth of autonomous AI agents
<https://mitsloan.mit.edu/ideas-made-to-matter/agentic-ai-explained>
substantially
expands this dynamic because they consume tokens and compute on their own
initiative.

The flat monthly subscription to an AI chatbot does not stop the meter from
running. The fixed cost may hide the meter from the consumer, but it does
not eliminate the never-ending metered incremental cost of every prompt,
every response, and every moment of machine deliberation. Users of flat-fee
chatbots rediscover the tollbooth when a usage limit is crossed, and the
model slows, shifts them to a less capable model, restricts use, or starts
adding incremental fees.1

Token economics favor organizations operating at scale. Large users can
negotiate volume or capacity commitments and employ engineers and
infrastructure to reduce inference costs through caching and routing. For
the small company competing with a bigger company, that can deploy more
intelligence per customer, that is a competitive disadvantage. For the
student competing with other students whose schools or parents can afford
more capable AI tutors, it is an educational disadvantage.

The token is the industry’s current denominator, and pricing schemes may
evolve to per task, per agent, or per outcome. Presently, there is an upsurge
in routers
<https://fortune.com/2026/08/09/why-every-company-wants-an-ai-model-router-right-now/>
to
direct requests to the least expensive model with adequate capability.
Undoubtedly, as the business evolves the billing unit may change as well.
But the logic will not. However intelligence is ultimately denominated, it
must be paid for. Regardless of whether the meter measures tokens, tasks,
agents, outcomes, or compute, the underlying economics remain.

A meter that never stops running will divide users by how much intelligence
they can afford, the quality of the intelligence they can access, and what
they are able to do with it.
Back to top
<https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid>
The token divide is more than the digital divide transplanted to AI

The term “digital divide” was widely popularized
<https://www.ntia.gov/page/falling-through-net-survey-have-nots-rural-and-urban-america>
in
1995 by Larry Irving
<https://www.internethalloffame.org/inductee/larry-irving/>, President Bill
Clinton’s assistant secretary of commerce for communications and
information. Early on, it was about access to and the affordability of a
computer and dial-up internet access. By the 2000s, it had become access
and affordability of high-speed broadband.

Broadband eventually spread, albeit imperfectly, slowly, and inequitably,
because network operators had a business model that rewarded expanding the
number of connections. Once sufficient broadband has been built, additional
usage can normally be bundled into a flat monthly price, and low-income
consumers are helped by subsidies.

The AI business model is different. Because every additional AI inference
requires additional computation, the AI business model wrestles with
absorbing, limiting, or passing along the additional cost incurred with
each new inquiry. It is a strain that becomes especially acute as
autonomous agents perform multistep tasks and consume tokens continuously.

The token divide is more than the digital divide by another name. Its most
significant difference is that unlike broadband, the target does not stand
still. Flat-price broadband access ultimately helped make the digital
divide manageable because a typical connection offered sufficient
throughput for most ordinary needs. AI may not have such a plateau. More
intelligence, or more capable intelligence, can enable qualitatively
different tasks and results, so even as the cost of a given level of
intelligence may fall, the frontier keeps advancing.

The meter keeps running while the frontier keeps moving.
Back to top
<https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid>
American institutions

Bridging the digital divide often fell to civic institutions.

Libraries became the on-ramp of internet access for all.  The local library
purchased a collection of computers, subscribed to an internet connection,
and opened the doors. That free access mattered, for instance, when job
applications began to be accepted only online.

Schools followed a similar path. In the early days of the internet, this
meant a connection to the computer lab down the hall. As computer literacy
and internet access became an increasingly important part of education,
connectivity moved out of the lab to each student’s desk.

Rural hospitals faced the challenge of implementing the advantages of the
internet in a situation where broadband did not always reach their
community.

Congress addressed these challenges by instructing the Federal
Communications Commission (FCC <https://www.fcc.gov/>) to create a universal
service funding <https://www.fcc.gov/general/universal-service-fund> structure
to subsidize school, library, rural hospital, and low-income access to the
internet. The subsidy worked because the economics cooperated as payments
to network providers helped them build, and payments to civic institutions
and low-income consumers helped them use.

But metered intelligence inverts the logic of those subsidies.

A library cannot pay for inference once given that every user session runs
the meter. The advantage of free use becomes the burden of continuously
recurring costs.

Schools face a similar challenge as intelligence becomes a per-student
usage cost. This can create a situation where one eighth grader has an
inexhaustible tutor while on the other side of town another student maxes
out token usage partway through an assignment.

The rural hospital now has connectivity but needs intelligence. As patient
volume and medical need drives AI usage, the quantity and quality of the
inference can vary based on what is affordable.

For public institutions, the token divide creates a recurring cost of
unknown size. The digital divide benefitted from how, once affordable
connectivity was in place, the marginal cost of the next web search, job
application, or homework assignment approached zero. The token divide has
no such advantage. Every use the institution seeks to enable now carries a
cost that grows to become an obstacle against that enablement.
Back to top
<https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid>
Good news, bad news

In 2025, large language model (LLM) “inference prices [fell] rapidly but
unequally across tasks,” according to
<https://epoch.ai/data-insights/llm-inference-price-trends> Epoch AI, a
nonprofit AI research institution. The rate of price declines on a
per-token basis ranged from nine to 900 times per year, the report found.

The Financial Times recently reported that
<https://www.ft.com/content/32a70a3c-7d28-40b4-808e-36edb58c7d01> “[l]eading
US AI labs such as OpenAI and Anthropic are releasing cheaper models as
they fight to retain cost-conscious customers who are switching to
cut-price alternatives from Chinese rivals.” On July 30, OpenAI slashed
<https://www.reuters.com/business/retail-consumer/openai-cuts-prices-smaller-models-businesses-scrutinize-ai-spend-2026-07-30/>
the
price of GPT 5.6 Luna—its fastest, cheapest model for high-volume tasks—by
80%, from $6 to $1.20 per million output tokens. The next day, Chinese AI
lab DeepSeek announced a new model with a rate card
<https://www.axios.com/2026/08/01/deepseek-model-cheap-ai-price-war> of 28
cents for the same one million output tokens.

Such absolute comparisons understate the complexity of understanding token
pricing. Because a token is a proprietary rather than a standardized unit,
each company’s model creates them differently. The result is that a million
tokens from one provider can represent meaningfully more or less than a
million tokens from another.

Putting aside the difference in tokens, falling prices do not solve the
token divide, and in fact, there are four ways that falling prices may
actually end up widening rather than closing the gap.
Usage effect

The 19th century British economist William Stanley Jevons identified a
behavior that resurfaces today in the token economy. Studying the consumption
of coal <https://archive.org/details/TheCoalQuestion>, Jevons identified
what economists call the Jevons paradox
<https://scienceinsights.org/what-is-the-jevons-paradox-and-why-does-it-matter/>:
When technology makes a resource cheaper, people end up using more of it.
Applied to tokens, it means that while per-token prices may fall, increased
usage drives total token spending upward. Lower unit costs invite more
uses, longer and more complex interactions, and increased autonomous
activity.
Frontier effect

The token gap is a moving target. Even amid price cuts, the powerful new
frontier models are typically priced at a premium, often with usage limits.
On July 30, OpenAI cut
<https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/>
its
lower tier by 80%, its middle tier by 20%, and its frontier model by zero.
Advantage accrues to those who can afford today’s frontier.
Relative-position effect

The measure of the token divide is the relation between the top and bottom
of capabilities. Everyone’s use of intelligence may rise because of lower
token costs, but the ratio of total intelligence deployed is still a matter
of who can pay. Falling prices can improve everyone’s absolute position
without eliminating relative inequality. Perhaps yesterday’s model is
sufficient for an eighth grader’s homework or the small firm’s customer
service, but that misses the point. The token divide is a competitive
condition. The student is measured against others. The firm bids against
rivals. The measurement of both is conducted based on the intelligence each
brings to the table. Broadband reached a plateau of sufficiency because
streaming a video at 100 megabits per second (Mbps) looks the same in a
mansion as it does in a trailer. Intelligence has no such plateau. When the
ceiling is constantly rising, “good enough” is constantly slipping behind
those who can follow the expanded intelligence upward.
Concentration effect

When token prices decline but the tollbooth simply moves to other layers of
the AI stack, cheap tokens and concentrated control arrive together.
Competition among model providers, cloud companies, and chip suppliers can
drive down the cost of intelligence. Yet, while there appears to be growing
competition among models, the infrastructure necessary to produce
intelligence is increasingly concentrated. A 2025 staff report
<https://www.ftc.gov/system/files/ftc_gov/pdf/p246201_aipartnerships6breport_redacted_0.pdf>
by
the Federal Trade Commission (FTC) identified how partnerships between the
largest cloud providers and the leading model developers could constrain
rivals’ access to compute and other inputs. The headlines about data center
construction foretell the AI pricing meter moving from models to the
computing power to run those models. The token divide thus becomes a
question of market structure, and closing it requires a policy that
promotes, preserves, and protects the competition that drives prices down.


Markets will continue to deliver abundant intelligence; whether they
deliver equitable intelligence is something falling prices cannot answer.
It is the reason the token divide will not fix itself.
Back to top
<https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid>
A division that never ends

As chairman of the FCC from 2013 to 2017, I oversaw the federal
government’s principal effort to bridge the digital divide. The solutions
Congress established and we applied were derivative of policy developed in
the 1930s to deal with electricity. The Rural Electrification Administration
<https://www.usda.gov/taxonomy/term/1811>, cooperative utilities
<https://en.wikipedia.org/wiki/Utility_cooperative>, and ultimately lifeline
rates <https://pubs.naruc.org/pub/AA519402-155D-0A36-3166-D7C97FCFD2C9> brought
power to every farm and household. Applying similar strategies
<https://www.fcc.gov/lifeline-consumers> helped attack the digital divide.

The four effects discussed above differentiate the token divide from the
earlier divides over electricity and the internet. They also demonstrate
why solutions originated in the 1930s do not resolve the 21st century
problem of access to the continually expanding capability of AI.

A household’s demand for electricity levels off once the lights,
refrigerator, and other applications are powered. The same holds true once
a certain level of network bandwidth is achieved. The quality of the
service being delivered was also constant across users. A kilowatt of power
was a kilowatt of power, no matter what it was used for, just as a bit of
throughput was a bit of throughput regardless of how it was used. Neither
of these conditions holds true for intelligence. Demand for intelligence
does not level off—it grows with each new capability and each new
application. A kilowatt may be a kilowatt in terms of what it delivers, but
all tokens are not created equal in their capability.

The digital divide had an identifiable finish line: an affordable
connection. The token divide has no such finish line because it is about
access to intelligence, its quality, and the relationship to the
intelligence available to others.

Here are three reasons that the solutions of the relatively quiet past are
not sufficient for the exponential future.
Capital expense vs. operating expense

The digital divide ended up being attacked by the application of capital.
New capital construction expanded the opportunity to connect across a wider
fiber and spectrum footprint. AI is also capital intensive, however,
expenditures for data centers and new models exacerbate the token divide by
expanding the capabilities of top-of-the-line AI.

For both the model owner and the user, the token divide is an operating
expense rather than a capital expense challenge. Society was able to build
its way through the digital divide. Once on the other side of the
connectivity challenge, the cost to access the next nonsubscription webpage
was invisible to the user. The token economy is the reverse: Access
consumes tokens and drives costs.
Moving target

Broadband policy relied on a benchmark of sufficiency. The FCC’s definition
expanded
<https://www.engadget.com/the-fcc-just-quadrupled-the-download-speed-required-to-market-internet-as-broadband-205950393.html>
from
downlink speeds of 25 Mbps in 2015 to 100 Mbps in 2024, a fourfold
increase. The scale of computation pushing the AI frontier is moving on a
radically different trajectory. Between 2010 and 2024, the computing power
<https://ourworldindata.org/data-insights/since-2010-the-training-computation-of-notable-ai-systems-has-doubled-every-six-months>
used
to train frontier models grew more than 100 million-fold, doubling roughly
every six months. This makes it difficult to develop a definition of
sufficiency in token access. At the exponential growth of AI, any benchmark
would be out of date almost immediately. Worse, a subsidy pegged to such a
baseline risks institutionalizing the divide by codifying yesterday’s
intelligence while the market moves on.
Quantity and quality

The ability to purchase throughput above 100 Mbps doesn’t necessarily
produce a smarter internet. The ability to purchase more AI tokens,
however, produces better outcomes, deeper research, more agents, and more
sophisticated reasoning.

Token access determines how often and for what AI can be used. The volume
of intelligence a user can afford can affect the value of the output.
Whether the user can reach the most capable models can determine the
quality of intelligence available to them.
Back to top
<https://www.brookings.edu/articles/the-ai-revolution-threatens-a-token-divide/?utm_campaign=Newsletters&utm_medium=email&utm_source=sendgrid>
What to do

A prescriptive “to do” checklist is inappropriate at this point in AI
history. This is, however, the right time to begin investigating causes and
solutions.

One thing is clear, however: What worked in the digital divide may be
inadequate for the token divide. Policymakers cannot conflate efforts to
deal with tomorrow’s token divide with remedies designed for yesterday’s
digital divide.

Such an investigation must produce a better understanding of the realities
of AI economics and their impact on consumers and competition. Congress
attempted to deal with this through Section 706
<https://uscode.house.gov/view.xhtml?req=(title:47%20section:1302%20edition:prelim)>
of
the Telecommunications Act of 1996 requiring
<https://legalknowledgebase.com/what-is-section-706-of-the-telecommunications-act-of-1996>
the
FCC to produce broadband deployment reports that created a factual record
for policy decisions. Congress can, and should, act again to develop an
understanding of the token divide through an appropriate body such as
the National
Institute of Standards and Technology <https://www.nist.gov/> or the National
Science Foundation. <https://www.nsf.gov/>

The path to token divide policy begins with measurement. Since the
tokenization
<https://www.mckinsey.com/featured-insights/mckinsey-explainers/what-is-tokenization>
process
differs from model to model, and different tokens have different
capabilities, there needs to be a unit of measure beyond simple token
counting such as cost per completed task or capability per dollar. Absolute
counts go stale almost immediately. A relational sufficiency standard
avoids that. Such a standard, indexed to a set percent of frontier
capability, would bring an equivalent to the cost-of-living indexing to the
token market.

Once the metrics are in place, it becomes possible to assess various
mitigations. Because new times challenge us to think anew, this becomes an
expansive exercise. Candidates for consideration could include but not be
limited to the following.
Institutional support

Support for schools, libraries, and rural health facilities was foundational
<https://www.fcc.gov/general/universal-service-schools-libraries> to
attacking the digital divide. We found that modernizing digital divide
programs was an ongoing challenge. We learned, for instance, that a program
for institutional support required empowering local institutions with the
leverage to deal with corporate power. When some local internet service
providers took advantage of government support to demand high rates from
schools, the FCC stepped up
<https://www.nytimes.com/2014/11/17/business/fcc-chief-aims-to-bolster-internet-for-schools.html>
to
give schools the competitive leverage of building their own last-mile
connection to the internet. The volume discount asymmetry of token pricing
invites similar expansive initiatives, including but not limited to
government buying consortiums and leveraging of government purchasing.
Open-weight safety

Open-weight models
<https://hai.stanford.edu/ai-definitions/what-is-an-open-weight-model> whose
core components are openly released offer low-cost access to intelligence.
These models frequently originate in China, raising national security
concerns. That by definition they are open to modification by anyone
further increases their risk. For the benefits of open-weight models to be
widely accessible, such problems must be addressed in a standardized
manner. A safety certification program that schools, hospitals, and other
institutions can adopt would open the door to a powerful tool to bridge the
token divide.
Public compute

A user can obtain an open-weight model for little to nothing. But “free”
weights do not mean free intelligence as the model still needs compute
power. Publicly supported compute can help break the chokehold of
concentrated and costly compute. Early efforts are worthy of support and
analysis. These include the National Science Foundation’s National
Artificial Intelligence Research Resource (NAIRR
<https://www.nsf.gov/focus-areas/ai/nairr>) pilot, California’s planned
CalCompute
<https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260SB53>,
and New York state’s Empire AI
<https://www.governor.ny.gov/news/governor-hochul-launches-empire-ai-consortium-make-new-york-global-leader-artificial>
.
Transparency

We have nutritional labeling
<https://www.fda.gov/food/nutrition-education-resources-materials/nutrition-facts-label>
for
our food, and the FCC requires broadband providers to post consumer labels
<https://www.fcc.gov/broadbandlabels>. Disclosure obligations would also
help AI users understand what they are getting as well as inform them when
they are being moved to a lesser model. When I asked for a report on the
tokens used in support of this paper, the information was unavailable to me
as a subscriber. It is such uncertainty that makes the token divide real
for librarians, school administrators, and businesses.
Nondiscrimination

Historically, owners of key infrastructure have used access to its
capabilities to control competition. The sine qua non—the essential,
non-negotiable condition—of access to both broadband and intelligence is a
policy of nondiscrimination to allow all who require the capability to be
able to use it on just and reasonable terms. When a frontier lab
manipulates its pricing, access rules, prioritization, and usage limits to
benefit its own or favored applications, user discrimination is at work
that widens the token divide. This does not have to mean common carrier
regulation, but so long as those who own key infrastructure are able to
constrain those capabilities for their own self-advantage, the AI
marketplace will be neither fully competitive nor functional.
Competition

Marketplace competition is at the heart of the world’s most successful
economy. But competition requires a supporting scaffold of rules. While it
may appear as though competition is driving down token pricing at the model
layer, the other necessary components of AI, from compute, to middleware,
distribution, and application integration, remain highly concentrated and
are growing more so. The AI marketplace is currently controlled by a
handful of corporate titans that are making their own rules for their own
benefit. Successful markets are markets with rules. Policies addressing the
integrated activities of the dominant AI firms would help to bridge the
token divide.
Democratizing AI

The internet democratized access to information. The digital divide taught
us, however, that technology does not democratize itself alone. Attacking
the digital divide was an effort to make that democratization a reality
through broad participation. Democratizing AI goes beyond that to address
whether access remains relevant as intelligence itself advances. The answer
to that question may depend on whether we prevent the emergence of a new
divide hidden inside the humble token.

Related Content
Why AI demand for energy will continue to increase
<https://www.brookings.edu/articles/why-ai-demand-for-energy-will-continue-to-increase/>
[image: A building complex that houses Compass Datacenters is partially
visible in Leesburg, Virginia, U.S., March 13, 2025.]
Energy Industry <https://www.brookings.edu/topics/energy-industry/>Why AI
demand for energy will continue to increase

Mark MacCarthy, David M. Klaus

August 12, 2025
Shaping the future of AI through responsible innovation
<https://www.brookings.edu/articles/shaping-the-future-of-ai-through-responsible-innovation/>
[image: A computer monitor with a parody of a tech company logo floats on
the waves while human hands reach up from the depths.]
Artificial Intelligence
<https://www.brookings.edu/topics/artificial-intelligence/>Shaping the
future of AI through responsible innovation

Nicol Turner Lee

September 23, 2025
AI makes the fight for net neutrality even more important
<https://www.brookings.edu/articles/ai-makes-the-fight-for-net-neutrality-even-more-important/>
[image: Supporters of Net Neutrality protest the FCC's then-recent decision
to repeal the program in Los Angeles, California, November 28, 2017. In
October 2024, the FCC approved a proposal to restore net neutrality rules.]
U.S. Government & Politics
<https://www.brookings.edu/topics/u-s-politics-government/>AI makes the
fight for net neutrality even more important

Tom Wheeler

April 9, 2024

Author
[image: Tom Wheeler, senior fellow. GS]
<https://www.brookings.edu/people/tom-wheeler/>
Tom Wheeler <https://www.brookings.edu/people/tom-wheeler/>Visiting Fellow
- Governance Studies
<https://www.brookings.edu/programs/governance-studies/>, Center for
Technology Innovation (CTI)
<https://www.brookings.edu/centers/center-for-technology-innovation/>
@tewheels <https://x.com/tewheels>

Glenn McKnight, MA
Virtual School of Internet Governance
Chief Information Officer
www.virtualsig.org
*YOUR SOURCE FOR INTERNET GOVERNANCE EDUCATION *
*Mobile  437-237-4655*
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://elists.isoc.org/pipermail/chapter-delegates/attachments/20260918/bd094f58/attachment-0001.htm>