THE PARKER EXPERIMENT

AI, operations, and the business of building things—tested in public, reported honestly.

Issue #8 | August 3, 2026 | Every token is a decision.

A quick note before we get into it.

This week I found out a Meta employee ran 280 billion tokens through an internal AI tool in a single month. That’s something like 2.3 million books’ worth of text, or roughly 50 books a minute, every minute, for 30 days straight. Somewhere else in that same week, Uber employees hit a 1,500 token cap because the company burned through its entire annual AI coding budget in about four months. Same technology, wildly different outcomes, and neither one is really a story about AI. Both are stories about whether anyone was watching the meter. I’ve spent 25 years managing budgets where the technology was never the real risk. The process around the technology was. Tokens are just the newest line item nobody’s built a system to watch yet. That’s Issue 8.

Now, the week.

LAYER 1 — THE PULSE

Tokens Went From Free to a Real Budget Line

Nufar Gaspar walked through the four stages of enterprise AI spending on The AI Daily Brief this week: oblivious, maximizing, anxious, and the target state, smart. Most businesses cycle through the first three before anyone notices, and the cycle usually ends with a bill that surprises somebody in finance.

The Meta and Uber examples above sit at opposite ends of that cycle. The top individual user on Meta’s internal leaderboard burned 280 billion tokens in a month. Uber capped employees at 1,500 tokens after running through its full 2026 AI budget in roughly four months and had to claw back control. Same underlying technology. The difference wasn’t the model. It was whether anyone had a system for watching what got spent and why.

There’s a second, quieter cost hiding in the first one. Anthropic shipped a new tokenizer on Opus 4.7 at the same sticker price, and independent analysis found real bills rose 12% to 27% anyway, because the same sentence now takes more tokens to say. Nobody announced a price increase. The math just changed underneath the price tag.

LAYER 2 — THE FEATURE

The Token Audit Any Small Business Can Run This Week

You don’t need a data team to find out where your AI spend is actually going.

Gaspar’s framework sorts every token a business spends into three buckets: tokens that teach, tokens that produce, and tokens that spin. Tokens that teach are the ones spent on genuine experimentation, a team member trying a new use case, building context, or figuring out what’s even possible. Tokens that produce are the ones spent generating an actual deliverable. Tokens that spin are the ones nobody would defend if you asked them to: idle agents left running, bloated context nobody trimmed, automations nobody remembered to turn off.

Gaspar’s own cautionary story made the case better than any slide could. She left an AI chief of staff agent running while traveling and came back to a $1,500 bill for an agent that produced almost nothing, 400 million tokens in and almost zero tokens out, a ratio near 3,000 to 1. Her fix, which she calls the weekend test, is worth stealing: if your AI bill keeps climbing on days nobody touched it, you have tokens that spin, and you can find them without hiring anyone.

“The most expensive token is the one that your best person is afraid to spend.”

— Nufar Gaspar, The AI Daily Brief, August 2

The metric that actually matters isn’t the price per token on a vendor’s page. It’s cost per accepted task, the total spend it takes to get a result you’d actually use. Databricks found this out directly: Sonnet 5 cost less per token than Opus 4.8, but needed more attempts to finish a coding task, so it ended up costing more overall, $2.09 versus $1.70. The cheaper model was the more expensive one to run. You can’t see that from a pricing page. You can only see it by tracking the task, not the token.

None of this requires a finance department. Pick one AI assisted workflow in your business this week. Track what it produces against what it costs, including the token drift you didn’t notice. If nobody could tell you that number today, that’s not a technology gap. That’s the same budget ownership gap I’ve watched businesses trip over for 25 years, just wearing a new unit of measurement.

AI PORTFOLIO WARS — WEEK 8 SCOREBOARD

LAYER 3 — THE SIGNAL

Quick hits from this week’s podcasts

• NVIDIA, Google, Microsoft, Meta, and eventually OpenAI signed an open letter defending open-weight AI models against a possible US ban, with Anthropic standing alone among major labs in refusing to sign. [The AI Daily Brief, July 28]

• More than 1,100 AI industry employees, including chief scientists at OpenAI, Anthropic, Google DeepMind, and Meta, asked the government to build tools that could deliberately slow AI development if it’s ever needed. [The AI Daily Brief, July 29]

• A hedge fund built on a correct thesis about AI still collapsed after running $120 billion in positions on $30 billion of capital, proof that being right about a trend and surviving it are two different skills. [The AI Daily Brief, July 31]

• Sam Altman told John Paul DeJoria that barbers and hairdressers are among the safest jobs from AI, since skilled trades are already in short supply and nobody wants a machine cutting their hair. [Success Story with Scott D. Clary, July 29]

• Amazon’s CEO said the company still won’t have enough capacity to meet AI demand through 2026 and 2027, with demand for 2028 already described as striking. [The AI Daily Brief, July 31]

BEFORE YOU GO

If you’re trying to figure out where AI actually fits in your business, that’s exactly what I help with at The Parker Group. No pitch, no pressure — just a straight conversation about what makes sense for your situation.

Book a free 30-minute call: parkergroup.us

Until next time,

Steve Parker

Founder, The Parker Group | AI Consultant, MBA, PMP

parkergroup.us | theparkergroup.substack.com