AI agents acquired more infrastructure than intelligence in the latest reporting window. Sandisk showed how sharply the price and mix of data storage moved. Amplitude reported growth while selling measurement for agent behaviour. Circle put transaction scale and new agent-payment claims beside the economics of reserve income. IonQ paired fast revenue growth with a much larger loss and a foundry acquisition.
This brief covers material filed or substantively announced between 5 August 2026 at 09:02 Iran time and 6 August 2026 at 09:02 Iran time (05:32 UTC to 05:32 UTC). Every company result below was accepted by the US Securities and Exchange Commission inside that window. Company descriptions of product adoption or causation remain management claims unless a reported metric directly supports them.
- Sandisk's quarterly revenue rose 51% sequentially, with management saying roughly two-thirds of that increase came from pricing.
- Amplitude's annual recurring revenue reached $410 million, while its GAAP operating loss widened to $35.2 million.
- Circle reported $73.3 billion of USDC in circulation and $14.8 trillion of quarterly onchain transaction volume.
- IonQ reported $80.1 million of revenue and a $1.868 billion GAAP net loss, making capital intensity impossible to ignore.
The common signal is operational, not promotional: an agent needs somewhere to retain state, a way to measure whether its actions worked, a controlled settlement path, and enough capital to support the underlying hardware. None of these filings proves a general AI investment thesis, and this article is not investment advice.
The 24-hour signal
| Layer | In-window evidence | Operational reading | Important limit |
|---|---|---|---|
| Storage | Sandisk quarterly datacentre revenue reached $2.977 billion | Memory and storage supply can become a major cost and capacity constraint | The filing does not isolate demand created by AI agents |
| Measurement | Amplitude ARR reached $410 million and free cash flow reached $23.7 million | Buyers are paying for behavioural analytics while agent observability becomes a product category | Product highlights are not audited adoption or outcome measures |
| Money | Circle reported $73.3 billion of USDC in circulation | Programmable settlement already operates at institutional scale | Reserve economics still depend heavily on rates and distribution costs |
| Frontier hardware | IonQ revenue reached $80.1 million as its adjusted EBITDA loss reached $120.3 million | Commercial demand and financing requirements can rise together | Quantum is adjacent infrastructure, not evidence of broad AI demand |
These are different ledgers. Storage revenue is not model revenue; annual recurring revenue is not cash; transaction volume is not profit; acquisition-funded capacity is not organic product performance. Our production AI cost-stack guide explains why operators should keep those units separate before approving a deployment.
Sandisk made storage the expensive part of the story
Sandisk's 5 August earnings exhibit reported fiscal fourth-quarter revenue of $8.965 billion, up 51% from the previous quarter. GAAP gross margin rose to 84.6%, and GAAP net income reached $6.903 billion. Management said roughly one-third of the sequential revenue increase came from higher volume and two-thirds from higher pricing.
Datacentre revenue was $2.977 billion, up 103% sequentially from $1.467 billion. For the full fiscal year, datacentre revenue reached $5.153 billion, up 437% from $960 million. Total fiscal-year revenue was $20.248 billion, up 175%. Sandisk also said it had signed five more New Business Model agreements and expanded its share-repurchase authorisation by $14 billion.
Those numbers support a narrow conclusion: high-value NAND demand and pricing moved dramatically. The underlying 8-K identifies the release as results for the quarter ended 3 July. It does not label datacentre sales as AI-agent revenue or disclose the workloads behind every shipment.
For an AI operator, the useful question is not whether storage is part of “the AI boom”. It is which state must be hot, which evidence must remain queryable, which traces can move to cheaper tiers, and how long each record must survive. Higher component pricing turns weak retention architecture into a direct operating cost. A system that stores every prompt forever can be both expensive and unnecessarily intrusive; our privacy-safe incident replay guide shows how to retain a minimal trace spine instead.
Amplitude sold measurement while the GAAP loss widened
Amplitude's 5 August earnings exhibit reported second-quarter revenue of $100.9 million, up 21% year on year, and annual recurring revenue of $410 million, up 22%. Remaining performance obligations reached $483 million, up 35%. Net cash from operations was $25.6 million and free cash flow was $23.7 million.
The same release reported a GAAP operating loss of $35.2 million, compared with $27.1 million a year earlier. Stock-based compensation and related payroll taxes were $27.4 million, while acquisition-related costs were $3.2 million. The filing therefore offers both growth and cost evidence; the non-GAAP result should not replace the GAAP one.
Management highlighted Global Agent, specialised monitoring agents, Model Context Protocol integrations, Agent Analytics and Wave. Those descriptions show where the product is moving: from retrospective dashboards toward continuous observation and suggested action. They do not disclose revenue for each AI feature, error rates, action volumes or customer outcome changes. The 8-K accepted at 20:38 UTC only establishes that the results release was furnished in-window.
The operating lesson is more durable than the product list. An agent needs a measurement contract before it acts: target metric, guardrails, baseline, attribution window, failure taxonomy and an owner who can challenge the recommendation. Instrumentation should preserve prompt, tool and policy versions without assuming that correlation proves the agent caused the result. The validation boundary in our safe-action gates guide is the relevant internal control: evidence and authorisation must be checked outside the model.
Circle connected agent payments to established [financial](/industries/finance) plumbing
Circle's 5 August results exhibit reported $73.3 billion of USDC in circulation at quarter-end, up 19% year on year. Quarterly onchain transaction volume was $14.8 trillion, up 151%. Total revenue and reserve income rose 7% to $701 million; revenue less distribution costs rose 15% to $289 million; adjusted EBITDA was $143 million.
Reserve income of $668 million rose 5% as average USDC circulation increased 25%, partly offset by a 66-basis-point decline in the reserve return rate. That relationship matters: network activity can expand while the yield on reserve assets falls. Circle's quarterly 10-Q, accepted later on 5 August, is the stronger source for accounting detail and risk factors.
Circle said its Agent Stack contained more than 900 paid services and that 99.3% of x402 agent-payment volume settled in USDC. This is a company-defined metric for a specified protocol activity, not a measure of all commerce conducted by agents. The release also described integrations or explorations involving BlackRock, BNY, DTCC and Standard Chartered, and said Arc's public mainnet was planned for 16 September. Future plans and partner exploration are not completed transaction evidence.
For finance teams, an agent payment is still a payment. Define the legal payer and payee, currency, approved counterparty, amount and velocity limits, sanctions and fraud checks, evidence retained, reversal or dispute route, and the person accountable for exceptions. A programmable rail can shorten settlement without deciding whether a purchase is authorised or economically sensible.
IonQ showed why revenue growth is not a capital shortcut
IonQ's 5 August earnings exhibit reported second-quarter revenue of $80.1 million, up 287% year on year and 20% above the midpoint of its prior guidance. The company raised full-year revenue guidance to $280 million–$290 million and reported $3.0 billion of cash, cash equivalents and investments at quarter-end.
Against that, GAAP net loss was $1.868 billion and adjusted EBITDA loss was $120.3 million. IonQ said the adjusted figure included $24.7 million of costs from its commercial relationship with SkyWater. Its acquisition of SkyWater closed on 31 July, after the reported quarter, and the new guidance excluded any SkyWater revenue contribution. The 8-K accepted at 20:05 UTC keeps those period boundaries visible.
Quantum computing is not a proxy for AI adoption. It belongs here only as an adjacent infrastructure and financing case: frontier hardware can report rapid revenue growth while consuming substantial capital and using acquisitions to secure manufacturing capability. Any enterprise evaluating hybrid quantum, high-performance computing or AI services should separate current paid workloads, research milestones, acquired capacity, pro-forma cash and forward guidance.
Build the agent stack as four controlled contracts
The filings suggest a practical architecture without requiring one vendor narrative:
- State contract: define what the agent may store, where it lives, how it is versioned and when it is deleted.
- Measurement contract: define success, guardrails, attribution, sampling and the evidence needed to explain a change.
- Settlement contract: define authority, limits, counterparties, screening, confirmation, reconciliation and recovery.
- Capital contract: distinguish recurring operating cost, committed capacity, acquisition effects and experimental spend.
Join the four with one run identifier and one accountable owner. A successful action should connect the input state and executable version to the measured outcome and any financial effect. A failed action should be replayable without retaining unrelated customer content. A payment should not proceed merely because the model predicts a favourable outcome.
Useful controls are concrete:
- tier storage by purpose and retrieval need rather than collecting everything;
- retain model, prompt, tool, policy and data versions for consequential runs;
- calculate product outcomes against a frozen baseline and declared window;
- require conventional policy code to enforce payment permissions and limits;
- reconcile transaction status with the source ledger rather than agent memory;
- review GAAP results, cash flow and acquisition effects beside growth metrics; and
- stop automation when required state, evidence or authority is missing.
Limits and what to watch next
This was an earnings-heavy window. Financial releases compress a completed quarter into selected metrics and management commentary; they are not real-time measures of the wider economy. Sandisk's price-driven growth may change with supply and contracts. Amplitude does not isolate AI-feature revenue or outcomes. Circle's agent-payment percentage covers a defined protocol slice. IonQ's post-quarter acquisition changes the future business but not the reported-period result.
The next useful evidence will be more disaggregated:
- Sandisk datacentre volume, price and customer-mix durability after the sharp fiscal-year move;
- Amplitude adoption and outcome measures for agent analytics, with GAAP cost and cash conversion;
- Circle's September Arc launch, live settlement activity, distribution economics and agent-payment definitions;
- IonQ's organic revenue, SkyWater contribution, cash use and foundry integration after closing; and
- any amended filing or regulator record that changes the facts above.
Agents do not become operational because the model can plan. They become operational when state is affordable and bounded, outcomes are measurable, money moves only under explicit authority, and the capital behind the system is visible.
