Token Maxxing vs Outcome Maxxing Is the Wrong Question

A live disagreement is running through Fortune 500 C-suites in public. In a public LinkedIn post, HubSpot CEO Yamini Rangan wrote that "outcome maxxing is greater than token maxxing." In a Wall Street Journal interview, Writer CEO May Habib described internal tokenmaxxing as essential for survival in the AI age. Appian CEO Matt Calkins compared the practice to the Soviet Union evaluating chandeliers by weight. Sequoia partner Sonya Huang told the same paper "we all should be tokenmaxxing."

Both camps include sitting CEOs and top-tier investors. Both camps are defensible. Many readers may recognize the argument already inside their own organizations.

This is where the dC/dt question gets sharper. AI capability is moving faster than executive operating policy can be revised on the normal cadence. The public debate has become binary: uniform tokenmaxxing that risks activity without outcomes, or uniform rationing that risks forfeiting a posture shift capital markets appear to be pricing. The binary is the error. The firms that seem to be pulling ahead are not choosing one doctrine. They are decomposing spend by workflow. Tokenmaxx where the workflow rewards it. Ration where it does not. Govern the portfolio as one.

What follows are suggestions of four workflows where tokenmaxxing may build position, the cadence a senior team can use to run the posture, and the mechanism by which outcome discipline and token discipline stop competing.

Observed and what follows from it

This argument relies on public statements from named executives, reported internal practices at AI and software firms, and early vendor product responses. It is an early operating read, not settled empirical proof.

The pro-posture case. At Nvidia's GTC conference in San Jose in March, CEO Jensen Huang said he would be alarmed if a $500,000 engineer did not consume at least $250,000 in tokens, and described a future in which every Nvidia engineer receives a token budget worth roughly half their base pay. Meta CTO Andrew Bosworth said in February that a top engineer spending the equivalent of their salary on tokens was producing 5x to 10x output, with no upper limit. According to a Fortune report, a Meta leaderboard called Claudeonomics surfaced 60 trillion tokens across more than 85,000 employees in 30 days, with the top individual averaging 281 billion, before it was shut down in early April 2026. OpenAI runs its own internal leaderboard; its top user consumed 210 billion tokens in a single week in March.

The posture is not confined to frontier AI labs. According to the Journal, Writer runs its own internal leaderboard; its March leader consumed nearly 11 billion tokens at roughly $50,000 per 10 billion. Sendbird has called its leaderboard highly impactful on internal adoption. Shopify has confirmed token use as one of several performance measures. Sequoia runs an internal leaderboard, and Sonya Huang has confirmed that many portfolio companies do the same.

Individual spend has moved into ranges that look more like capital allocation than tool usage. According to a New York Times report, a single Claude Code user billed more than $150,000 in a month, and Mechanize co-founder Ege Erdil described continuous agentic workflows running at 700 million tokens per week per agent.

The skeptical case. Rangan argued that what AI builds has to maximize outcomes for customers and for the company, and that the unit of analysis is specific workflows in teams or functions that can dramatically improve productivity. Jellyfish CEO Andrew Lau told the Journal that a company could tokenmaxx all day and produce outcomes that were not what was desired. Blitzy CEO Brian Elliott compared the metric to measuring sales team revenue by cold calls made. An OpenAI employee told the New York Times the practice "doesn't seem sustainable." Earlier this year, Salesforce announced a new metric called Agentic Work Units, which translates tokens and compute into completed work, explicitly as a rebuttal to tokenmaxxing.

What follows. The stronger inference is that the two camps are closer than the binary suggests. Rangan's post points to specific workflows as the unit of analysis. Habib's defense rests on per-action ROI being the wrong level. They are approaching the same operating principle from opposite sides. The token bill alone cannot tell the executive team whether the firm is building position or producing theater. The aggregate outcome metric alone cannot tell the executive team where to allocate tokens. Decomposition by workflow is what lets a firm pursue outcomes at the portfolio level and tokenmaxx at the workflow level without contradiction.

None of this yet proves that a high-token operating posture generalizes across every industry or function. It shows that the pressure to adopt one is already arriving before the measurement systems are mature.

One additional signal before the four cases. The Claudeonomics leaderboard was pulled after its data surfaced externally, raising concerns that aggregate token patterns could reveal strategic priorities, staffing, and R&D allocation to outside observers. Token spend data is strategic information. The instrument needs to treat it that way.

The CFO tension. Token spend looks like opex on the invoice and closer to venture capital on the balance sheet. A single engineer billing $150,000 in a month is not a typical SaaS line. It looks more like an underwriting decision priced as an operating expense, against a return function still being developed. The OpenAI employee's concern about sustainability lands in this frame, and Rangan, Lau, and Elliott have made the public case that the return function may not materialize as described. The tension does not defeat the strategic case. It defines what the instrument has to do.

Early signal, not proof

The evidence base is narrow by industry. Frontier labs, hyperscaler infrastructure, enterprise AI, commerce platforms, mid-market SaaS, and venture capital. An F500 industrial, insurer, or consumer brand executive has a fair question. Does this generalize.

The useful answer is that the question is being answered by the market in parallel to any internal operating plan. Investors are reportedly asking portfolio companies about token posture in diligence. Engineering candidates are asking about token budgets in offers. Enterprise customers are beginning to ask vendors what AI configuration produces the service they are paying for. Software engineering commentator Gergely Orosz has written that inside large tech companies, it is becoming a career risk to not use AI at an accelerated pace regardless of output quality. Salesforce has built a product metric to rebut the posture, which suggests the posture is mainstream enough that the largest enterprise software vendor in the world has built a counter-product.

For many firms, the practical decision is no longer whether to engage at all, but whether to engage on their own terms or react to the market's.

This is the dC/dt shape of the question. The posture appears to be spreading through capital markets, talent markets, customer conversations, and vendor roadmaps faster than most F500 operating plans can revise to address it. Leading the conversation internally preserves position. Waiting absorbs the lag.

Threading the disagreement

Rangan is right that outcomes matter more than tokens. Habib is right that waiting for per-action ROI risks forfeiting the posture. Outcome discipline applies at the portfolio. Token discipline applies at the workflow. Name which workflow earns which, and the binary dissolves.

The four workflows are the operating form of that thread.

The four workflows that earn the burn

1. Capability discovery

When a new model, reasoning capability, or workflow category enters the organization, the firm does not yet know what the tool can do. Spending aggressively during this phase produces the map of what works and where. That map becomes the input to every subsequent build-or-buy decision.

Anecdotally, the firms that have extracted the most value from frontier AI over the past eighteen months appear to have spent aggressively during discovery and rationed aggressively at scale. The sequencing is the strategy. Discovery has an outcome, which is the map, so Rangan's logic and Habib's logic can both hold here.

2. Option value on frontier capability

Some decisions reward seeing everything the tools can see. A litigation strategy with nine-figure exposure. An M&A thesis with a narrow window. A drug discovery triage call. A threat model for critical infrastructure. Running these at full context, maximum reasoning effort, and across multiple frontier models is the cost of not missing what the firm could have caught.

The asymmetry favors the buyer. If the lower-cost configuration produces the same answer, the firm spent modest margin. If the higher-cost configuration catches something the lower one missed, the firm holds the decision. Huang's $250,000 figure is a version of this logic applied to engineering labor. For consequential decisions in an F500 portfolio, the expected-value logic is at least plausible enough to justify deliberate testing.

3. Competitive intelligence through use

Firms that run broadly across models and at scale tend to develop proprietary knowledge that narrower firms do not. Which models handle which tasks. How providers behave under capacity pressure. Where the frontier actually is. When new capabilities shift what is possible.

A meaningful part of this knowledge is produced only by direct, repeated use inside the firm's own workflows. Consultants do not see those workflows. Benchmarks are often stale on arrival. The firms building this capability now appear to be developing an information advantage that will take time to close. The tokens are the investment. The intelligence is the asset. The Claudeonomics episode made visible that this asset carries enough strategic weight that the organizations producing it are already thinking about how to protect it.

4. Speed of iteration

Token discipline has a latency cost. Every review gate is time a competitor may be using to ship. For workflows where speed is the advantage, the token bill is the price of velocity.

Bosworth's reported read is that his best engineer, at full token posture, is operating at roughly 5x to 10x rationed productivity. That figure is a company claim, not an audited measurement. Lau and Elliott's point that activity can be mistaken for output has force here, and the right response is to pair velocity workflows with the kind of outcome measurement Salesforce is building toward with Agentic Work Units. Velocity is the right posture when the firm can measure what the velocity produced. Without that measurement, it is harder to distinguish velocity from theater.

This lands hardest in product development, customer-facing iteration, and response to competitor moves. The firm that reaches a working version in days, even at higher per-unit cost, preserves position. Inference costs have been falling underneath the decision, not rising against it. Paying for velocity today is paying less for it tomorrow, provided the outcome measurement tells the firm velocity was the right call.

Cadence

A senior team can build the posture in ninety days and revisit it each quarter.

• Within thirty days, produce the first decomposition. Token spend by workflow, tagged against the four strategic categories and a fifth for production workloads where rationing is correct. The first version will be imperfect. The discipline is producing it.

• Within sixty days, the CEO, CFO, and Chief Strategy Officer align on which workflows earn which posture for the next two quarters. The outputs are not budgets. They are positions, owned by named executives. The failure mode to watch for is misclassification. A firm that tags a production workload as discovery, or insurance as velocity, will burn capital on the wrong curve and call it strategy. The decomposition is only as good as the category discipline behind it, and that discipline is built by challenge, not by assertion.

• Within ninety days, the senior team runs the instrument as a live document, reviewed monthly, adjusted as the models and the market move. Each category carries an outcome metric alongside its token allocation. That is where Rangan's discipline enters the instrument and Salesforce's product direction becomes useful rather than rhetorical. This is the cadence of a strategic variable, not a cost line.

The firms that pull ahead are likely to be the ones where the CEO and senior team know which workflows deserve the burn, which do not, and can defend the answer against Rangan's skepticism and Habib's urgency at the same time. Capability, market, and competitive posture are all changing faster than the cycle most operating plans were built to address. Closing that gap is the work.

The only question that matters is whether your executive team has a position, or a number.