AI's Unit-Economics Test Reaches Hardware, Software and Payments

This release cycle makes the economics of AI unusually visible from the chip supplier to the software seat and the payment rail. AMD reported the compute capacity being sold; HubSpot described agents monetized through seats and credits; Duolingo showed an AI feature being paced against delivery cost; and Circle attached agent commerce to a settlement asset and protocol. ISM and the U.S. Treasury then supplied the operating and funding backdrop. Five records landed inside this edition's rolling 30-hour window. AMD's after-market release arrived about 33 hours before the Tehran cutoff, so the window was extended within the permitted 48 hours to capture the latest global earnings cycle. ZharfAI's thesis is narrower than “AI is booming”: the market is beginning to expose unit economics, and every layer must now prove that greater usage can survive its own cost structure.

AMD provides the hardware anchor. Second-quarter revenue rose 50% year over year to $11.536 billion, while Data Center revenue reached a record $6.7 billion, up 107% and equal to roughly 58% of company revenue. GAAP gross margin was 54% and GAAP operating income was $1.990 billion. The comparison needs care: the year-earlier quarter included about $800 million of inventory and related charges caused by U.S. export controls on Instinct MI308 products. AMD's company-defined non-GAAP gross margin was 56%, and its third-quarter outlook of approximately $13 billion in revenue, plus or minus $300 million, with about 56% non-GAAP gross margin is a forecast rather than a result. The Data Center segment also combines EPYC server CPUs, Instinct accelerators, and other products; the release does not isolate accelerator revenue. Demand is confirmed at segment scale, but neither product mix nor customer return is fully visible.

HubSpot moves the test from compute supply to enterprise distribution. Second-quarter revenue rose 20% as reported and 17% in constant currency to $911.7 million. GAAP operating income was $43.3 million, a 4.8% margin, compared with a $24.6 million loss and negative 3.2% margin a year earlier. Company-defined non-GAAP operating income was much higher at $185.3 million, or 20.3%, so the accounting basis matters. Operating cash flow was $222.8 million and non-GAAP free cash flow was $167.9 million. Customer count grew 14% to 306,446, while average subscription revenue per customer increased 4% to $11,800. Management said its AI agents are monetized through core seats and credits and emphasized predictable pricing. Yet the release does not separately disclose agent revenue, credit consumption, or audited customer return. The quarter shows a distribution and monetization mechanism, not proof that every deployed agent pays back.

Duolingo offers the clearest feature-cost experiment. Daily active users increased 23% to 58.7 million, paid subscribers rose 17% to 12.7 million, revenue grew 18% to $298.5 million, and bookings advanced 8% to $289.1 million. GAAP gross margin improved 20 basis points to 72.6%. The company attributed the upside versus its roughly 71% expectation partly to a measured rollout of AI-powered features and AI cost efficiencies. It also said most new Super subscribers now receive access to Video Call, which it plans to expand. That is useful evidence of staged deployment: experience, adoption, and inference expense are being managed together. It is not an independent learning-outcomes study. Net income fell 26% to $33.2 million, adjusted EBITDA margin declined 530 basis points to 25.9%, and non-GAAP free-cash-flow margin fell 790 basis points to 26.3%. Stable gross margin therefore coexists with weaker profit and cash margins elsewhere in the model.

Circle connects the software unit to a financial rail. The company said Circle Payments Network's Agent Stack gives AI agents access to more than 900 paid services and reported that 99.3% of payment volume using the x402 agent-payment protocol settles in USDC. The boundary around that number is essential: it is Circle's metric for x402 volume, not evidence that USDC handles 99.3% of all agent payments, and the release does not disclose the dollar value represented by that share. Even so, the mechanism is concrete. An agent can discover a priced service, pay through a machine-readable protocol, and settle with a stable-value token. The next proof is not the number of catalog entries but recurring paid usage, service reliability, fraud and dispute handling, regulatory compliance, and net economics after payment and infrastructure costs.

Circle's quarter also shows why volume and financial value cannot be conflated. USDC in circulation was $73.3 billion at quarter-end, up 19% year over year, and onchain volume reached $14.8 trillion, up 151%. Revenue and reserve income increased only 7% to $701 million. Reserve income rose 5% to $668 million because 25% growth in average USDC circulation was partly offset by a 66-basis-point decline in reserve return rate. Distribution, transaction, and other costs were $412 million, up 1%, while adjusted operating expenses rose 23% to $146 million. Net income of $48 million improved by $530 million, but the prior-year comparison carried IPO-related stock compensation and related charges; company-defined adjusted EBITDA was $143 million, up 8%. The durable equation is circulation multiplied by reserve yield, less partner distribution and operating cost—not headline transaction volume alone.

The July ISM Services survey is the operating counterweight. The Services PMI registered 54.1, business activity 59.1, and new orders 57.2, all signaling expansion through a diffusion index. Employment remained in contraction at 47.4, however, while the Prices Index rose to 70.3. Prices have exceeded 60 for 20 consecutive months; 17 industries reported paying higher prices and none reported decreases. Respondents identified higher prices for software licensing and support, memory products, and computers, while some described shortages of technical labor and memory. One anonymous respondent linked a small staffing reduction with AI implementation. A single survey comment cannot establish that AI caused the broader employment reading, and a diffusion index does not measure output or margin. The stronger conclusion is that service demand remains firm while input and hiring conditions can still squeeze an AI deployment's payback.

Treasury's quarterly refunding is not an AI event, but it keeps the capital hurdle in view. The department offered $125 billion of securities to refund approximately $96.3 billion maturing on 15 August and raise about $28.7 billion of new cash. It expects to maintain current nominal coupon and floating-rate-note auction sizes for at least the next several quarters. Its financing assumptions include a $950 billion end-September cash balance and a possible late-October peak near $1.05 trillion, plus or minus $50 billion; planned buybacks reach up to $38 billion for liquidity support and $25 billion for cash management over the quarter. These figures do not mechanically set a software company's borrowing cost. Auction demand, term yields, credit spreads, and each issuer's risk do. But infrastructure and product investment still compete against a visible sovereign benchmark for capital.

Across the six records, “unit economics” means different denominators and they should not be collapsed. AMD's unit is capacity sold inside a mixed segment; HubSpot's is a customer, seat, or credit; Duolingo's is an active learner and AI interaction; Circle's is circulation, settlement, or paid service; ISM measures the breadth of business responses; Treasury measures financing supply. GAAP results, company-defined adjusted measures, forward guidance, and network statistics answer different questions. The shared test is conversion: can added usage produce reported revenue, can reported revenue preserve gross margin, can gross profit cover distribution and operating expenses, and can the resulting cash return clear the cost and risk of capital? This framework is less spectacular than a model benchmark, but it is much harder to imitate with promotional language.

The next disclosures can make the thesis falsifiable. Watch AMD's Data Center mix and GAAP margin after the export-control comparison normalizes. At HubSpot, look for agent-specific revenue, credit consumption, retention, and customer outcomes rather than adoption language alone. At Duolingo, compare broader Video Call availability with engagement, learning evidence, inference cost, and total operating leverage. Circle needs to disclose actual x402 payment value, repeat usage, take rates, failures, compliance cost, and how declining reserve yields interact with circulation growth. ISM's Employment and Prices indexes will show whether services expansion becomes easier to staff and cheaper to deliver, while Treasury auctions will reveal the market's reception of supply. For operators, including Iranian teams facing currency and cloud-access constraints, the practical discipline is to meter cost per successful outcome, fallback, latency, human review, and payment failure before scaling. This is an operating framework, not investment advice.