Why Alex Karp's tokenmaxxing warning applies to your ad platform

On July 1, 2026, Alex Karp put a useful phrase around a billing problem that had already become familiar inside enterprise AI budgets: “there is a reason why those selling tokens refuse to charge based on value.” He was talking about tokenmaxxing, the habit of treating more AI tokens consumed as if it automatically meant more work accomplished, more productivity created, or more business value delivered.[1]

The reason Karp’s tokenmaxxing criticism matters beyond enterprise AI is that the same measurement trap already runs through automated media buying. Tokens are to enterprise AI what spend volume, impression volume, clicks, and blended platform ROAS are to PMax, Advantage+, and AI Max when the same vendor controls delivery, attribution, and the victory screen.

Palantir’s own nine-point AI sovereignty manifesto sharpened the charge with the line that “tokenmaxxing hijacks your value orientation.”[2] That is a better description of bad AI measurement than most dashboards manage. It names the moment when an input metric becomes emotionally and politically easier to defend than the thing the buyer actually wanted.

There is a caveat worth getting out of the way early. Palantir has a commercial stake in making token-based AI consumption look reckless and making value-tracked, enterprise-controlled AI look responsible. Critical coverage of the manifesto has noted that self-interest.[3] That does not make the measurement problem fake. It means the framework needs to be separated from the vendor selling it.

The token counter is not the value counter

The tokenmaxxing examples are excessive enough that they almost read like parody. Business Insider reported internal Meta data showing 60.2 trillion tokens consumed across the company in a 30-day period, with the top single user consuming 281 billion tokens.[4] Fortune reported that Uber had burned through its entire 2026 AI token budget by April.[5] TechCrunch cited Jellyfish data from 7,548 engineers showing measured throughput roughly doubling while token cost rose tenfold.[6]

Those facts should be handled carefully. The Meta and Uber numbers are reported internal metrics, not independently auditable public ledgers. The Jellyfish dataset is a specific customer population, not a law of software productivity. Still, they describe a recognizable curve: more machine activity, more billable consumption, unclear marginal value.

That curve is not exotic to anyone who has scaled an automated ad campaign past its efficient range. The first budget increase finds demand the account was missing. The second widens reach. The third starts buying softer intent, remarketing spillover, brand-adjacent searches, cheaper attributed conversions, or inventory that makes the platform dashboard look busy while finance waits for revenue that does not arrive with the same slope.

The issue is not that tokens, impressions, clicks, or spend are useless. They are real operating inputs. A model cannot answer without tokens. A campaign cannot acquire customers without media exposure. The problem starts when the vendor reports the input with the confidence of an output, then prices the relationship around the part it can expand.

How the same metric trap shows up in automated media buying

In enterprise AI, token consumption can be made to look like productivity. In paid media, spend absorption can be made to look like scale, impression delivery can be made to look like market coverage, and blended ROAS can be made to look like profitable growth. Each metric contains some information. None of them, by itself, answers the buyer’s real question: what did this system add that would not have happened anyway?

This is why the tokenmaxxing debate matters so directly to bidding. It is a budget allocation problem before it is a branding problem. The system has a spendable input, a vendor-controlled optimization loop, and a reporting layer that can reward the machine for finding more measurable activity rather than more incremental business.

A PMax campaign that absorbs another $20,000 and reports more conversions has not necessarily created $20,000 worth of new demand. It may have captured demand already created by brand, email, affiliates, organic search, offline sales activity, or prior paid exposure. Advantage+ may find cheaper attributed purchases without finding better customers. AI Max may broaden query coverage while quietly changing the mix of intent. The dashboard can be directionally useful and still be structurally conflicted.

The vendor incentive is simple. A token seller benefits when more tokens are consumed. An ad platform benefits when more budget clears through the auction. Both can also create genuine value. That is precisely why the measurement standard has to be stricter, not looser. The presence of real value in some cases is what makes blended reporting so dangerous in the others.

Blended ROAS is the media buyer’s token counter

Blended platform ROAS is not a fake number. It is worse than that: it is a number that can be true inside the platform’s rules and still answer the wrong question. If an automated campaign reports a 5x ROAS after absorbing demand that would have converted through brand search or direct traffic, the platform did record revenue against media exposure. It did not prove the business gained five dollars for every new dollar spent.

The same distinction matters in tokenmaxxing. An engineering team may generate more code, more summaries, more test drafts, more tickets, or more internal responses. Some of that output may be useful. Some may create review burden. Some may replace work that would have happened anyway. Some may move cost from writing to verification. Counting tokens does not settle any of that.

For media buyers, the equivalent failure happens when an AI campaign is judged on volume-weighted averages instead of marginal contribution. The platform reports total conversions. The buyer needs incremental conversions. The platform reports average CPA. The buyer needs marginal CPA at the next tranche of spend. The platform reports attributed revenue. The buyer needs revenue that finance can recognize after returns, cancellations, discounts, contribution margin, and customer quality are accounted for.

Where the dashboard gets permission to overstate

Automated campaign reporting tends to overstate when three conditions appear together. First, the platform has broad discretion over where budget goes. Second, attribution gives credit for touchpoints near existing demand. Third, success is judged inside the same interface that spent the money.

• Budget discretion lets the system move from high-intent demand into cheaper or broader inventory without making the trade-off obvious.
• Attribution proximity lets the system claim conversions that were already likely to occur.
• Blended reporting hides whether the last dollars spent performed worse than the first dollars.
• Opaque auction and placement data make it difficult to separate new reach from recycled demand.

Tokenmaxxing has the same permission structure. A team is given access to a powerful AI system. Usage rises. Vendor bills rise. Internal dashboards show activity. Unless the organization connects that activity to shipped work, reduced cycle time, lower support cost, better decisions, or revenue, the usage layer becomes the proof layer by default.

What to demand instead

Palantir reportedly built an internal tool Karp called “demastibatory” to track token value per dollar rather than token volume for its own sake.[4] Set aside the name. The useful part is the independent verification layer. The buyer needs a system that asks what each unit of consumption produced after cost, baseline, and quality are applied.

For paid media, that standard is not complicated in principle. It is just harder than accepting the default dashboard.

• Use incrementality where the account can support it: geo tests, conversion lift studies, audience holdouts, or clean pre/post designs when true holdouts are unavailable.
• Track marginal ROAS by spend tier, not only blended ROAS across the full campaign.
• Reconcile platform revenue to finance-recognized revenue after refunds, cancellations, discounts, and delayed order adjustments.
• Separate new customers from returning customers, and measure whether automated campaigns are changing customer mix or simply harvesting the cheapest attributed buyers.
• Audit query, placement, creative, audience, and product-feed mix wherever the platform still exposes enough data to identify where scale is coming from.

The most useful media-buying report is often less glamorous than the platform summary. It shows what happened when spend moved from one band to the next. It identifies whether the campaign’s incremental revenue flattened while attributed revenue kept climbing. It flags whether a bid strategy improved acquisition or merely changed the attribution mix. It gives the person paying the invoice a reason to increase, cap, restructure, or cut budget that does not depend on the vendor’s own definition of success.

A simple version can live outside the ad platform: daily spend by campaign type, platform-reported conversions, backend orders, new-customer share, gross margin, refund-adjusted revenue, and a baseline estimate for what would likely have happened without the extra spend. It will be imperfect. It will still be better than treating the platform’s activity count as the business outcome.

The buyer’s line

The cleanest question to ask any AI system is not whether it did more. It is what the next unit of consumption produced that the business would not have received otherwise.

For token-based AI, that means value per dollar of tokens after review cost and actual workflow impact. For automated media, it means incremental revenue, marginal ROAS, customer quality, and finance-verified outcomes after the baseline demand is removed. A vendor can still earn more budget under that standard. It just has to earn it with value, not volume.

Do not accept platform activity metrics as value metrics. Do not let AI ad systems define success with numbers they can inflate by consuming more of your budget.

References

1. Alex Karp's July 1 CNBC interview, CNBC, 2026-07-01.
2. Palantir's 9-point manifesto, Business Insider, 2026-07.
3. Critical coverage of Palantir's AI sovereignty manifesto, TNW.
4. Alex Karp Compares Tokenmaxxing to a Porn Addiction, Business Insider, 2026-06.
5. Tokenmaxxing is over, Fortune, 2026-05-28.
6. Tokenmaxxing is making developers less productive than they think, TechCrunch, 2026-04-17.