Dear Readers,

A note from us: University students receive our Saturday Deepdive for free when they register with their university email address at: https://getsuperintel.com/plus-whitelist

Welcome to Intelligence from the Community, our Sunday format where a selected author from the Superintelligence community publishes an original essay or analysis. The idea hasn't changed: some of you are researchers, some are operators, some are engineers building the systems everyone else writes about. That expertise deserves space.

This week's piece comes from Amish Regmi, an AI engineer at Klaviyo who previously built inference infrastructure and agentic systems at Amazon. Amish tackles something that has been bugging me for months: the way "AI is exponential" gets thrown around as if it were a single, self-evident fact. It rarely comes with the numbers that would make it testable. What is the base of the exponent? What is the doubling time? Which curve are we even talking about? Amish goes through the data, separates confirmed steep exponents from fast hillclimbs and broken instruments, and arrives at a conclusion that is more useful than the slogan: the transition will be governed by mismatched slopes.

If you think you have an original contribution for this series, apply here: https://docs.google.com/forms/d/e/1FAIpQLScjSo4iYH24p5-p-PdPCcVoSayJRhEamhBOp_Srt1Jb9rI4zA/viewform?pli=1

All the best,

Kim Isenberg

In Today’s Issue:

🤑 Google I/O turns Gemini into an agent stack

💵 KPMG brings Claude into enterprise delivery

📈 SynthID becomes a wider AI watermarking layer

📉 AI data centers hit local opposition

✨ And more AI goodness…

AI is broadly associated with exponential growth, but these statements are rarely backed by specifics: the base of the exponent, the unit of time over which the multiplier is measured, the number of sustained periods, and the doubling time.

A curve that doubles every two months is a different growth pattern from one that doubles every two years. Growth per training run, per model generation, per calendar year, and per dollar of inference spend are different measurements and should not be treated as interchangeable.

A precise exponential claim should be stated as a multiplier per unit interval. For example, y(t)=y0 b^(t/tau), where b is the base multiplier and tau is the time unit. The doubling time is T2 = tau log(2) / log(b). Moore's Law is precise because it names the doubling time: transistor density roughly doubles every two years.

2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn what’s changing and how to prepare.

Download the Toolkit

The cost to match a fixed AI capability tier has been declining by at least an order of magnitude per year. Even this framing is conservative. Epoch AI's broader benchmark set is steeper. Their analysis shows the median inference-price decline is roughly 50x/year, with task-level rates ranging from 9x to 900x/year.

The a16z framing of "LLMflation" is useful because it separates unit-price collapse from aggregate spend. They tracked that the cost to match a comparable MMLU score fell from about $60 per million tokens to about $0.06 per million tokens.

This paradox holds: unit prices can collapse while aggregate AI spend rises. That happens because usage expands, agentic workflows multiply the number of calls, enterprises shift workloads onto newer and more expensive frontier models, and budgets often include fixed capacity commitments rather than pure per-token API spend.

One critical technical implication sits underneath inference pricing. Training compute has long been the focus, but a significant portion of compute is shifting toward test time. "Train-to-Test" scaling work shows that modern labs are jointly optimizing model size, training tokens, and inference samples.

Anthropic is the clearest public revenue exponent. The company reported a $9B annualized revenue run rate at the end of 2025. That figure grew to more than $14B in February 2026 and surpassed $30B by April 6, 2026. For context, OpenAI stated on April 1, 2026, that it was generating $2B per month.

The surrounding markers reinforce the curve. Anthropic raised a $30B Series G at a $380B post-money valuation. Claude Code, the company's coding-focused product, launched publicly in May 2025 and exceeded a $2.5B run rate by February 2026. On the demand side, more than 1,000 business customers now spend over $1M per year with Anthropic.

3. Enterprise generative AI spending

Enterprise generative AI spending went from $1.7B in 2023 to $11.5B in 2024 and $37B in 2025. The average is steep, but the multiplier is already shrinkingg: 6.8x, then 3.2x. That makes the curve an early-stage hillclimb rather than a confirmed sustained exponent.

4. The broad market and small-enterprise adoption

Gartner forecasts $2.528T in worldwide AI spending in 2026. Adoption is broad but not yet deeply integrated. Goldman Sachs reports that 76% of surveyed small businesses use AI, but only 14% say it is fully embedded in core operations. Capital is following the same pattern of concentration, with the OECD reporting that AI firms received 61% of global venture-capital investment in 2025.

Infrastructure investment is massive, but it is not the same curve as demand. Epoch AI's filings-based analysis finds that hyperscaler capex grew about 72% per year from Q2 2023 through Q4 2025. If that trend continues, it implies roughly $770B in 2026 capex. This is the central tension: exponential demand curves meeting a slower, capital-intensive supply curve.

Capability is the curve that would justify the rest of the stack. It is also the curve we measure least reliably. OpenAI's Frontier Evals team deprecated SWE-bench Verified on February 23, 2026, after finding that at least 59.4% of audited hard failures had flawed tests.

Even infrastructure configuration can move scores. Long context windows are another measurement trap. "Lost in the Middle" research shows that models can underuse information placed in the middle of long inputs.

Furthermore, "Chain-of-Thought" should not be treated as a faithful audit log. Anthropic reports that reasoning models often do not reveal the factors that truly influenced their answers.

Mathematical reasoning is producing real signals, but not yet a clean curve. Scientific American reports that AI tools have helped move about 100 Erdős problems into the solved column since October 2025.

Users do not want a chatbot. The convergence point is one audio-native, vision-native AI that sees what the user sees, hears what the user hears, and acts across tools. Meta launched Muse Spark on April 8, 2026. OpenAI's GPT-4o announcement framed the same product direction. Each modality expansion multiplies token volume per interaction, feeding the revenue exponent and justifying infrastructure investment.

Those requests came from non-traded BDC investors in Q1 2026, and most got back roughly half of what they asked for. Moody's U.S. BDC sector outlook: Negative.

On Percent's marketplace that same quarter: new issuances, scheduled payments, 0.44% lifetime net loss rate on asset-based deals since inception.† The difference is structural: concentrated corporate loans with redemption windows that close at manager discretion vs. asset-based finance with 6–24 month deal terms. 14.6% net ABS returns LTM after losses (3/31/26).† Starting at $500.

See how private credit works when it's built for individual investors.

_{Alternative investments are speculative. No assurance can be given that investors will receive a return of their capital. †Past performance is not indicative of future results. Terms apply.}

The broad claims that AI is exponential are too vague to be useful. A more rigorous framing is narrower and more actionable: some AI curves are confirmed steep exponents, some are fast hillclimbs, and some cannot yet be measured reliably.

Inference price decline and Anthropic’s revenue growth are the clearest high-slope curves. Enterprise spending, small-business adoption, and hyperscaler capex are rising quickly, but their slopes and accounting scopes differ. Capability measurement remains the weak link: software benchmarks are contaminated, long-context retrieval is position-sensitive, and chain-of-thought is not a faithful audit trail.

The practical conclusion is that the transition will be governed by mismatched slopes. Unit costs can fall while aggregate bills rise. Revenue can compound faster than infrastructure. Adoption can be broad while integration remains shallow. Capability may be improving, but current measurement instruments are unreliable and sometimes misleading. Some AI growth metrics exhibit exponential behavior, but without specifying the curve, time unit, doubling time, and sustained duration, the phrase “AI is exponential” is not analytically useful.

Amish Regmi is an AI engineer at Klaviyo, where he works on AI agents for data analytics. He previously worked at Amazon on AI infrastructure and agentic systems. He is primarily interested in multi-agent system misalignment evaluations and the economic signals around frontier AI.

Sources