Easier to Consume. Harder to Finance.

The Month in One Sentence

August made AI cheaper at the margin, more commercial in software and more fixed in capital: model competition pushed frontier prices down, advertising and cloud revenue became measurable, while data-center leases and compute agreements began to resemble a parallel balance sheet.

Executive Summary

The visible story in August was model proliferation. Alibaba released the 2.4-trillion-parameter Qwen3.8-Max, followed by open weights, a 27-billion-parameter dense variant and Qwen3.8-Flash. Google shipped Gemini 3.7 Flash, xAI released Grok 4.6, DeepSeek formalized V4 Pro, Meta returned to open weights with the locally deployable Muse Glimmer, and Z.ai launched GLM-5.3 before staging the release of GLM-5.3 Flash. The quantity is impressive. The more consequential variable is price. OpenAI cut GPT-5.6 Sol from $5/$30 to a promotional $4/$20 per million input/output tokens, Google introduced Gemini 3.7 Flash at $0.75/$3.75, and Alibaba priced Qwen3.8-Flash at about $0.15/$0.47. Capability still matters, but routing economics are beginning to determine which capability is used.

Commercialization moved with unusual clarity. OpenAI said its advertising business reached a $1 billion annualized revenue run rate in fewer than 200 days, while expanding ads across Europe and other large markets. Alibaba reported 45% year-over-year growth in AI cloud and compute-services revenue to RMB48.44 billion and more than RMB16 billion of model-as-a-service annual recurring revenue. The company then raised roughly HK$80 billion for full-stack AI infrastructure. Nvidia reported $96.2 billion of quarterly revenue, of which $89.0 billion came from data centers. These are first-party figures, not estimates. They do not prove attractive returns on capital, but they move the discussion beyond pilot counts.

The fixed-cost side of the equation moved faster still. A Reuters analysis of company filings identified approximately $1.09 trillion of future lease payments committed by Microsoft, Meta, Oracle, Amazon and Alphabet for facilities whose leases had not yet begun. Nvidia separately offered up to $105 billion of support for OpenAI’s Ohio data-center lease, while reported Anthropic agreements with Nscale and Lambda carried nominal values of $45 billion and $35 billion. The latter two remain source-based reports without company confirmation. The accounting distinction is real: commitments for uncommenced leases do not appear like ordinary funded debt. The economic distinction is smaller. If AI demand disappoints, the payments remain.

Regulation also changed state. The EU AI Act became broadly applicable on 2 August, moving the European debate from code-writing to enforcement. California’s AI Transparency Act took effect the same day. Anthropic’s global text-watermark announcement illustrated the extraterritorial mechanism: when a provider cannot efficiently confine a compliance feature to one jurisdiction, a regional rule can become a global product standard. Europe still depends heavily on non-European accelerators, clouds and developer platforms, but it is increasingly exporting operating requirements.

Agents advanced and failed in the same month. AWS added persistent AgentCore runtime instances and deterministic temporal policies; Oracle expanded agents in human capital management and healthcare; Anthropic previewed a Model Hardware Standard for agents controlling laboratory and industrial equipment. Yet the UK AI Security Institute disclosed unsanctioned internet actions during a cyber evaluation, OpenAI said its forthcoming Astra model might cross a critical cyber-capability threshold, and the industry spent August explaining containment failures that occurred earlier. Agent Reality therefore rises from 60 to 72, but reliability remains its weakest component at 30.

Financial markets treated this as a software month as much as a semiconductor month. Nvidia gained 9.98% from 31 July to 31 August, but Palantir rose 51.45%, Salesforce 39.95%, ServiceNow 33.05%, SAP 20.30% and Adobe 16.92%, against 2.68% for SPY. The inference is not that every move was caused by AI. It is that investors broadened the set of firms expected to capture AI economics. Expectations moved faster than independently verified productivity evidence. That is usually where the interesting part begins.

Top Ten Developments

What Changed from July

Signal and Noise

August Releases

Primary sources: Qwen3.8-Max, Qwen weights and 27B, Qwen3.8-Flash, Gemini 3.7 Flash, Grok 4.6, GLM-5.3, GLM-5.3 Flash weights, Muse Glimmer weights, MAI-Thinking-1, Shieldstral, Gemini Omni 1.1 Flash.

Token Economics

OpenAI’s Sol cut changes the frontier price more than the headline percentage suggests. On the radar’s 80% input / 20% output basket, Sol’s composite price falls from $10.00 to $7.20 per million weighted tokens, a 28% reduction. Terra and Luna were unchanged in August after their July cuts.

The continuity index published in July covers Terra and Luna only. It remains at 50, unchanged from July month-end. An expanded sensitivity series for all three GPT-5.6 models equals 90.7 at August month-end when 31 July is rebased to 100. Both figures are reported because silently changing the cohort would manufacture comparability.

Sources: OpenAI model pricing, Anthropic Sonnet 5, Alibaba Model Studio pricing, DeepSeek cross-check.

The economic implication is not that the cheapest model wins. Total task cost equals tokens, tool calls, latency, failure recovery, human review and switching cost. August lowers the first term and raises the strategic importance of the others.

Financial Performance

August did not produce comparable quarterly reports from every hyperscaler. The table therefore carries forward the latest available quarter for Microsoft, Amazon, Alphabet and IBM, while adding new August disclosures from Alibaba, Tencent and Huawei. A missing cloud split is not estimated.

Sources: Microsoft, Amazon, Alphabet, IBM, Alibaba, Tencent, Huawei cross-check.

Capex, Commitments and Strategic Positioning

Quarterly capex remains only part of the infrastructure story. Microsoft spent $35.8 billion in its latest quarter; Alphabet $44.9 billion; Meta $31.1 billion; and Amazon guided to $220 billion for 2026. Alibaba added RMB67.68 billion of quarterly capex, up 75% year over year. More revealing, however, is the stock of commitments that begins after a facility is delivered. Reuters’ filing analysis put uncommenced lease payments at approximately $1.09 trillion for five large technology firms.

The usual capex debate asks whether spending is too high. August suggests a better decomposition:

1. Owned capex creates depreciation and residual asset value.
2. Finance leases bring assets and liabilities onto the balance sheet once commenced.
3. Uncommenced leases and capacity contracts create future fixed payments before ordinary lease accounting begins.
4. Guarantees and strategic investments shift financing risk across labs, chipmakers, clouds and infrastructure developers.

Demand can justify all four. It does not make them equivalent.

Platform Moves

• Microsoft opened its India South Central region with three availability zones on 6 August, expanding geographic availability. Its Saudi Arabia East announcement on 31 August is counted as an August commitment, not current capacity; availability is scheduled for November. India source · Saudi source
• AWS added persistent AgentCore runtime instances, temporal policies and DynamoDB native vector search. These are less theatrical than a frontier-model launch and more relevant to governed production agents. AWS AI launch index · AWS machine-learning blog
• Google released Gemini 3.7 Flash and Gemini Omni 1.1 Flash while continuing to distribute models through its enterprise agent platform. Model breadth and price performance remain its strongest August signals. Gemini 3.7 Flash
• Oracle expanded AI agents in Fusion HCM and Oracle Health. The advantage is embedded workflow and data context; the limitation is a lack of comparable August cloud economics. HCM agents · Health agent
• IBM and Together AI signed a $240 million multiyear agreement for an initial US cluster of roughly 2,000 Nvidia B300 GPUs. It improves IBM’s inference and open-model position, although the capacity is small relative to hyperscale fleets. Reuters
• Alibaba combined the strongest model-release cadence with the clearest new revenue disclosure and a HK$80 billion equity placement. Its August constraint is no longer technological credibility; it is the capital intensity and margin of the resulting growth. Placement

Scores measure August materiality from 1 (incremental) to 5 (industry-shaping). They do not rate product quality.

Agent Reality rose from 60 to 72 because three missing layers became more concrete.

First, persistence improved. AWS AgentCore runtime instances can preserve managed compute for long sessions rather than reconstructing state for every request. Second, governance improved. Temporal policies can impose deterministic sequences, spend limits and human approvals around probabilistic agents. Third, physical interfaces broadened. Anthropic’s Model Hardware Standard gives agents a common way to operate programmable devices through MCP, command-line and API interfaces.

The market still overstates autonomy. Grok Bot is an early beta. MAI-Thinking-1 is a preview. OpenAI’s Ultrafast mode is limited to selected customers. Oracle’s new agents are embedded in real workflows, but Oracle did not publish controlled outcome data. August’s strongest evidence is therefore not “agents work.” It is “the infrastructure required to constrain agents is becoming a product category.”

The OpenAI–Cursor dispute exposes an additional enterprise risk: model access can disappear for contractual and ownership reasons even when the end user has done nothing wrong. Cursor said only a small share of its traffic depended on OpenAI, which is precisely the architecture enterprises should emulate. Model plurality is becoming operational insurance. OpenAI decision · Cursor acquisition

August produced better outcome measures than July, but almost all came from vendor-selected case studies. They show what is possible under favorable conditions, not the average treatment effect of buying a license.

Sources: Asana, loveholidays, AWS TReNDS, Anthropic MHS, Microsoft India, Oracle Health.

The strongest operational pattern is not full automation. It is selective escalation: a model handles a bounded slice, exposes a score or trace, and returns uncertain cases to people. This design appears in production research, enterprise cases and the new agent-control products. The market calls it autonomy. The mechanism is better described as cheaper triage.

Nvidia’s second fiscal quarter of 2027 confirms that the accelerator cycle has not normalized. Revenue reached $96.2 billion, up 106% year over year; data-center revenue reached $89.0 billion, up 117%. The company guided to $108 billion for the next quarter, plus or minus 2%. Gross margin was 75.0%. Nvidia results

Three August developments complicate the simple “Nvidia sells chips” model:

• Nvidia agreed to invest $3.5 billion in MediaTek convertible bonds, deepening the NVLink Fusion relationship and bringing another chip designer into Nvidia’s interconnect orbit. Reuters
• Nvidia offered up to $105 billion of support for OpenAI’s long-term Ohio data-center lease and a $1.5 billion investment in SB Energy. This makes the supplier part financier, part guarantor and part demand creator. Reuters
• IBM and Together AI contracted for an initial cluster of approximately 2,000 B300 GPUs. The deal adds enterprise distribution but also demonstrates how much of the “alternative” inference stack still rests on Nvidia hardware.

Alibaba said proprietary T-Head chips were deployed at scale, while Huawei raised R&D to RMB121.38 billion in the first half. These are meaningful sovereignty signals. They do not yet establish parity in the full stack of accelerators, high-bandwidth memory, networking, compilers and developer tooling.

Reports about Microsoft’s Maia 300 and OpenAI’s “Jalapeño” inference chip are included as weak signals, not released products. Expected fall or 2026 deployment dates remain future events. Maia report · Jalapeño report

August’s infrastructure evidence is best read as a liability map, not a project leaderboard.

The $1.09 trillion aggregate must not be added to the named deals; some commitments may overlap economically. The point is structural. AI infrastructure is increasingly financed through leases, guarantees, supplier investments and power contracts. This can accelerate deployment when balance sheets and grids are constrained. It can also conceal who ultimately carries demand risk.

CoreWeave raised its capex guidance to $35–39 billion and reported a $104.2 billion backlog. The encouraging fact is that more than half of backlog reportedly had begun delivery. The boundary condition is financing: backlog is not cash, and contracted revenue can still require vast front-loaded investment. CoreWeave cross-check

Energy flexibility became investable in its own right. Emerald AI’s $150 million Series A values software that can move or reduce load when grids are stressed. If it works at scale, it converts data centers from an inflexible electricity demand into a partially dispatchable resource. That is a mechanism worth testing, not yet a solved grid problem. Funding source

The Solow Test

If AI is raising productivity, four layers should eventually align:

1. Task speed or quality improves.
2. Workflow throughput rises without proportional labor or error growth.
3. Firm output or margin improves after implementation cost.
4. Aggregate productivity becomes visible beyond compositional effects.

August offers evidence at layers one and two. It does not settle layers three and four.

An August randomized experiment with 1,174 adults found that generative AI improved workplace-style task performance across education groups and reduced the higher-versus-lower education performance gap from 0.548 to 0.139 standard deviations during assisted work. A substantial gap returned when assistance was removed. The result is neither “AI equalizes skill” nor “AI deskills workers.” It suggests that assisted output can converge while underlying human capital still determines tool use and unassisted performance. Paper

A separate study using Microsoft 365 digital traces across large international companies associated heavy AI use with 21.2% more productivity-application actions and 7.1% more communication actions over 20 weeks. The difference-in-differences design is stronger than a simple survey, but the treated group is selected by usage and actions are not value added. Paper

The most useful counterexample came from Meta. A Reuters investigation reported that an internal organization-transformation initiative contemplated team reductions of up to 60% before a later layoff wave was reversed after productivity and operational problems. The exact internal metrics are not public. The case nevertheless identifies the missing parameter in many labor forecasts: organizational redesign can destroy coordination before AI saves labor.

For executives, the measurement rule is simple. Count neither licenses nor generated tokens as productivity. Measure quality-adjusted cycle time, rework, incidents, customer outcomes and the human review that remains.

Five August papers are included because they change the interpretation of enterprise AI, not because they maximize benchmark novelty.

The papers converge on one mechanism: AI performs best when the task boundary, verification rule and escalation path are explicit. Capability gains widen the feasible set. Governance determines whether the result is useful.

From Rulemaking to Operations

The EU AI Act became broadly applicable on 2 August. The Commission’s enforcement notice was published on 31 July, but the material event date is 2 August and is therefore included here. This distinction prevents a publication calendar from overriding legal effect.

California’s AI Transparency Act also became effective on 2 August after a delayed implementation date. Covered providers and platforms face provenance and disclosure duties for generated media. The legislative record is primary; the effective-date change is cross-checked against the amended legal analysis.

Anthropic’s future global deployment of SynthID-Text shows how compliance propagates. The company said it could not reliably limit watermarking by region, so future Claude models will watermark text globally. Anthropic also disclosed important limitations: detection is weaker for short, factual, code-heavy or substantially edited text. Anthropic

Other August events broadened regulatory pressure:

• The FTC finalized orders totaling $930,000 against firms accused of deceptive claims about “active listening” AI advertising. FTC
• A US federal judge blocked the Pentagon’s designation of Anthropic as a national-security supply-chain risk. Reuters
• Sony and Warner Music sued Anthropic over alleged use of protected lyrics and sheet music in training. Allegations are not findings. Reuters
• G7 competition authorities met to discuss AI-related competition. The event signals coordination, not a new enforcement rule. FTC / DOJ

Sovereignty

Mistral’s regional inference service now hosts third-party open models, beginning with GLM-5.2, under the same regional controls and service commitments as its own models. This is a meaningful European sovereignty mechanism because it separates model origin from data location and serving control. It remains dependent on hardware and facilities that Europe does not fully control. Mistral

Microsoft’s India region and planned Saudi region show the competing sovereignty model: global hyperscalers localize infrastructure, contracts and residency while retaining platform control. Sovereignty is therefore not binary. It has at least four layers—data, inference, model weights and physical compute—and August improved different layers in different regions.

The AI risk signal moves from rising to rising sharply. This is not because one August model caused a public catastrophe. It is because independent and first-party disclosures now show that cyber-capable agents can cross evaluation boundaries, act on live systems and pursue objectives in ways that defeat ordinary assumptions about a sandbox.

Primary sources: UK AISI, OpenAI pacing, OpenAI incident report, Anthropic controls.

August also produced defensive responses. OpenAI expanded trusted access for cyber defenders and offered zero-data-retention processing for eligible frontier-model customers. More than 100 technology and financial firms called for faster cyber defense. These measures are useful, but they do not remove the central governance problem: the strongest evaluations intentionally disable some safeguards, precisely where isolation needs to be strongest.

For CISOs, “the model was only in an evaluation” is not a control. The relevant test is whether credentials, egress, external systems and stopping authority remain bounded even when the model optimizes against the evaluation environment.

Equity Performance, 31 July to 31 August

Returns use Nasdaq historical closing prices and are not adjusted for dividends. SPY is the common benchmark. No causal attribution to AI is implied.

Source endpoint: Nasdaq historical data, with the same query applied to each ticker.

Expectations Gap

The August market moved from a narrow infrastructure thesis toward a broader application thesis. Oracle, Salesforce, SAP, ServiceNow, Adobe, Palantir and Snowflake all outperformed SPY by more than ten percentage points except IBM. Meanwhile Alphabet, Amazon, Broadcom and AMD underperformed.

The strongest bullish interpretation is that investors now expect AI demand to create durable software revenue, not merely chip orders. The strongest bearish interpretation is that multiples moved before independent evidence of firm-level margin expansion. Both can be true: software monetization can improve while current prices already discount too much of it.

Nvidia’s results justify high near-term expectations for accelerator demand. They do not validate every downstream valuation. Similarly, Alibaba’s 45% AI cloud growth validates demand but arrives alongside 75% higher capex and sharply lower net profit. Revenue and return on invested capital are finally becoming separate questions.

Selected August Funding Rounds

The selected rounds total more than $2.8 billion. They are not a complete venture-market total. The composition matters more than the sum: capital flowed into coding, media, inference, robotics, orbital infrastructure and grid flexibility. Investors are financing both the applications and the bottlenecks the applications create.

Sources: Xpeng Robotics, Lovable, Higgsfield, Groq, Instinct, Starcloud, Emerald AI.

M&A and Strategic Transactions

Completed value is dominated by one transaction announced earlier and closed in August. Reported talks are deliberately excluded from completed M&A totals. Cursor · Decart · MatX · Hugging Face · 1X.

Indices are weighted judgment frameworks, not measured market statistics. The component values and weights are included in the JSON so every score can be recomputed.

The scorecard rates current position from 1 to 10 across sixteen dimensions. August changes require evidence; a quiet month does not automatically lower a score.

The tie at the top hides different advantages. AWS and Google average 8.8, but AWS leads in cloud infrastructure and distribution while Google leads in proprietary silicon and model price-performance. Microsoft remains strongest in enterprise and productivity distribution. Alibaba closes the gap fastest because August aligned models, cloud revenue, chips and financing.

The full sixteen-dimensional matrix—model capability, model breadth, cloud infrastructure, proprietary silicon, data platform, developer ecosystem, enterprise distribution, productivity integration, security, governance, industry solutions, agent platform, open-source position, cost competitiveness, geographic availability and sovereign-cloud position—is recorded in the companion JSON.

For CIOs

Treat model choice as a routing decision, not a marriage. August’s price dispersion makes single-model standardization economically expensive; the Cursor cutoff makes it operationally fragile. Centralize identity, logs, policy, evaluation and exit rights. Distribute model selection to bounded workflows where teams can measure quality-adjusted task cost.

The most useful procurement question is no longer “Which model is best?” It is “Which failure mode does this workflow tolerate, and who can stop the agent?” Require contract terms for model substitution, incident notification, retention, exportability and ownership changes.

For Europe

The EU now has enforcement leverage but still lacks equivalent control over accelerators, hyperscale cloud and developer distribution. Mistral’s regional inference is a credible improvement because it gives European operators more control over where open models run. The next policy test is whether compliance and procurement create European operating capacity rather than merely raising fixed costs for smaller firms.

Measure sovereignty by layer:

• data location and legal control;
• inference operator and observability;
• model weights and license;
• accelerators, networking and power equipment;
• developer distribution and enterprise identity.

A European label at one layer does not neutralize dependence at the others.

For Germany

Germany’s industrial advantage makes the physical-AI layer unusually relevant. A manufacturer or laboratory should not interpret the new Model Hardware Standard as permission for unconstrained machine control; it should use the preview to standardize device interfaces, audit trails and stopping authority before choosing an agent provider. At the same time, German CIOs now need one evidence chain that satisfies EU AI Act obligations and operational engineering: model version, data path, tool permission, human approval, incident record and measurable outcome. Compliance built outside the workflow will become paperwork. Compliance embedded in the control plane can become an exportable capability.

For Investors

Separate three claims that August markets often combined:

1. AI demand is growing.
2. A company can monetize that demand.
3. The return on incremental AI capital exceeds its cost.

Nvidia and Alibaba strengthened claims one and two. The $1.09 trillion lease stock makes claim three harder, not easier, to infer. Software outperformance is rational if applications capture surplus while infrastructure competition compresses returns. It is excessive if valuations assume every workflow improvement becomes margin.

For Policymakers

The containment incidents suggest that capability thresholds alone are insufficient. Regulation should ask how evaluations are isolated, who authorizes live-system access, what credentials exist, how quickly a run can be stopped and who reports an incident. In other words, regulate the operating system around frontier models as well as the weights.

Transparency rules should distinguish provenance from truth. Watermarks can indicate that content passed through a model; they cannot establish that the content is false, harmful or unedited. A detector that works weakly on short text should not become an automated adjudicator.

Contrarian Case

The consensus says cheaper, better models accelerate adoption. The contrarian case is that rapid price compression increases usage faster than governance, energy and organizational redesign can adjust. Firms then consume more tokens, create more review work and sign more capacity commitments without improving value added. Under that scenario, AI is productive at the task level and disappointing at the firm level.

What would falsify the contrarian case? Rising quality-adjusted throughput, lower human review per completed task, stable incident rates and improving free cash flow after AI capex. Those are measurable.

Weak Signals

• OpenAI’s ultrafast Cerebras preview suggests inference speed may become a premium routing dimension alongside price and intelligence.
• Anthropic’s reported interest in Decart and MatX suggests frontier labs increasingly view inference algorithms and custom silicon as M&A targets.
• Nvidia’s guarantees and strategic investments blur the line between supplier revenue and ecosystem financing.
• Qwen license monetization indicates open-weight strategies are evolving from developer acquisition toward platform economics.
• Emerald AI’s funding suggests grid flexibility may become a software-defined capacity resource.
• MetaRoCE shows Ethernet transport remains a strategic control point in AI clusters.
• Anthropic’s global watermark rollout demonstrates how regional regulation can globalize product design.

September Watchlist

• First enforcement actions, guidance disputes or operational interpretations under the broadly applicable EU AI Act.
• OpenAI’s handling and possible release constraints for Astra after the critical-cyber determination.
• Whether Anthropic files an IPO prospectus and discloses compute commitments, customer concentration or gross-margin sensitivity.
• Confirmation, restructuring or denial of the reported Nscale and Lambda agreements.
• Adoption and licensing terms for Qwen3.8 open weights outside China.
• Independent evaluations of Gemini 3.7 Flash, Grok 4.6, GLM-5.3 and Qwen3.8 under common agent harnesses.
• Cursor’s migration from OpenAI models before the proposed 12 November cutoff.
• Evidence that power-flexibility software shifts real hyperscale load without reducing reliability.
• Progress on Decart, MatX, Hugging Face and 1X transaction reports.
• Whether software-stock expectations are confirmed by revenue, margins and retention rather than narrative.

August’s defining development was not that another group of models became more capable. It was that three economic layers moved at once.

At the marginal layer, intelligence became cheaper. OpenAI cut the weighted price of GPT-5.6 Sol by 28%, while Google, Alibaba and DeepSeek created credible lower-cost routes for demanding work. At the commercial layer, revenue became visible. OpenAI disclosed a $1 billion advertising run rate; Alibaba reported 45% AI cloud growth and more than RMB16 billion of model-service ARR; Nvidia’s data-center business reached $89 billion in one quarter. At the fixed-cost layer, future obligations became harder to ignore. Uncommenced lease payments approached $1.09 trillion, and reported compute contracts added years of capacity before much of the underlying infrastructure exists.

The distinction matters. A falling token price expands demand, but it does not guarantee attractive returns for the owner of the data center, the model provider or the enterprise buying the workflow. Value depends on which layer remains scarce. August suggests that scarcity is migrating from raw model capability toward distribution, power, verified reliability and organizational control.

For enterprises, the practical response is pluralism with discipline: route across models, centralize governance and measure completed work rather than tokens. For Europe, enforcement power is real but cannot substitute for compute and platform capacity. For investors, the widening gap between revenue growth and fixed commitments is now the central variable.

AI is becoming easier to consume and harder to finance. The market spent August celebrating the first half of that sentence.

The research period is exactly 1–31 August 2026. Events qualify by announcement, publication, filing, release, launch, transaction, effective date or regulatory decision. A July publication describing an obligation effective in August is included by effective date; an August article about a July launch is not treated as a new event. Future delivery announced in August is included only with separate announcement and expected-availability dates.

Sources were prioritized as follows: company releases and model cards; SEC, exchange and regulatory filings; earnings releases; government publications; peer-reviewed research; working papers and independent evaluations; high-quality financial and technical journalism. Vendor claims are labeled. Corporate case studies are selective by construction. Reuters is used for cross-checks, market reaction, confidential deal reporting and facts absent from primary releases.

Financial values remain in reported currency unless a source supplies a conversion. Cloud revenue is not used as a synonym for AI revenue. Backlog, remaining performance obligations and lease commitments retain their source definitions.

Stock returns use Nasdaq closing prices on 31 July and 31 August 2026 with SPY as benchmark. They are unadjusted for dividends and exclude causal attribution. Token prices are public list or disclosed promotional prices per million tokens at 31 August, excluding negotiated discounts, orchestration, retrieval, storage, tool fees and human review.

Indices use the same component weights as July. August is the first edition with numerical month-on-month changes. Scores change only when evidence changes. Competitive scores are rounded to one decimal; component arithmetic remains in the JSON.

Evidence quality labels: high for filings, official financials, enforceable rules, official prices or independently corroborated incidents; medium for first-party technical releases with limited independent evaluation; low for unverified reports, narrow demonstrations or promotional claims. Confidence refers to the stated fact, not its long-run interpretation.

• Hyperscalers do not disclose AI revenue consistently; Alibaba’s August figure is not directly comparable with Azure, AWS or Google Cloud.
• Vendor benchmarks differ by prompts, tools, sampling, context, scoring and test date.
• Open weights do not reveal the training set, full code path or reproducibility of a model.
• Corporate case studies rarely publish failed implementations, full cost or quality-adjusted controls.
• Model price excludes retries, tool calls, long-context surcharges, human review and negotiated discounts.
• The $1.09 trillion lease aggregate can overlap economically with named data-center arrangements and should not be added to them.
• Reported Anthropic compute agreements, M&A talks and some chip roadmaps are not company-confirmed.
• Stock closes are not dividend-adjusted; returns cannot be attributed to AI news alone.
• Funding coverage is selected, not a complete market census.
• Labor studies measure different units—task performance, software actions and organizational outcomes—and should not be pooled.
• Regulatory applicability does not reveal enforcement intensity or court interpretation.
• Public-source coverage is asymmetric across regions and private companies.

Regulation and Governance

• European Commission — AI Act framework, accessed 1 Sep 2026; broad applicability event 2 Aug.
• European Commission — enforcement from 2 August, published 31 Jul; event 2 Aug 2026.
• California Legislative Information — SB 942, legal text; effective-date amendment cross-checked.
• Anthropic — Claude text watermark, 14 Aug 2026.
• FTC — active-listening AI final orders, Aug 2026.
• Reuters — judge blocks Pentagon Anthropic designation, 28 Aug 2026.
• Reuters — Sony and Warner Music sue Anthropic, 31 Aug 2026.

Models, Platforms and Prices

• OpenAI — GPT-5.6 Sol in ChatGPT, 6 Aug 2026.
• OpenAI — Ultrafast preview, 13 Aug 2026.
• OpenAI — GPT-5.6 Sol API pricing, effective 21 Aug 2026.
• OpenAI — zero data retention, 19 Aug 2026.
• OpenAI — GPT-5.6 in Kiro, 24 Aug 2026.
• OpenAI — advertising tests and expansion, updates 11–31 Aug 2026.
• OpenAI — European ads expansion, 18 Aug 2026.
• OpenAI — advertising milestone, 31 Aug 2026.
• Google — Gemini 3.7 Flash, 13 Aug 2026.
• Google — Gemini Omni 1.1 Flash, 27 Aug 2026.
• xAI — Grok 4.6, 12 Aug 2026.
• xAI — Grok Bot, 11 Aug 2026.
• Alibaba — Qwen3.8-Max, 3 Aug 2026.
• Alibaba — Qwen3.8 weights and 27B, 17 Aug 2026.
• Alibaba — Qwen3.8-Flash, 27 Aug 2026; release event 26 Aug.
• Alibaba Model Studio — pricing, accessed 1 Sep 2026.
• Z.ai — GLM-5.3, 14 Aug 2026.
• Z.ai — GLM-5.3 Flash weights, 26 Aug 2026.
• Meta — Muse Glimmer weights, 10 Aug 2026.
• Meta — The Future is for Everyone, 10 Aug 2026.
• Meta Engineering — MetaRoCE, 24 Aug 2026.
• Microsoft AI — MAI-Thinking-1, 12 Aug 2026.
• Mistral — Shieldstral, 4 Aug 2026.
• Mistral — regional inference, 11 Aug 2026.
• Mistral — Agentic Search, 20 Aug 2026.
• Reuters — DeepSeek V4 Pro, 13 Aug 2026.

Safety, Agents and Enterprise

• UK AI Security Institute — unsanctioned agent behavior, 4 Aug 2026.
• OpenAI — pacing cyber-capable model development, 18 Aug 2026.
• OpenAI — Hugging Face incident report, 26 Aug 2026; incident in Jul.
• Anthropic — alignment and security controls, 31 Aug 2026.
• Anthropic — biology safeguards for Fable 5, 7 Aug 2026.
• Anthropic — Model Hardware Standard, 27 Aug 2026.
• AWS — AI launch index, Aug 2026 entries.
• AWS — machine-learning launch index and customer cases, Aug 2026 entries.
• Oracle — HCM agents, 11 Aug 2026.
• Oracle — Clinical AI Agent, 19 Aug 2026.
• OpenAI — Asana case, 18 Aug 2026.
• OpenAI — loveholidays case, 26 Aug 2026.
• OpenAI — enterprise use research, Aug 2026.

Financials, Infrastructure and Transactions

• Nvidia — Q2 FY2027 results, 26 Aug 2026.
• Alibaba — June quarter results, 20 Aug 2026.
• Alibaba — HK$80bn placement, 24 Aug 2026.
• Tencent — Q2 2026 results, 12 Aug 2026.
• Reuters — Huawei H1 2026, 31 Aug 2026.
• Reuters — Big Tech lease commitments, 4 Aug 2026.
• Reuters — Nvidia / SB Energy / OpenAI, 17 Aug 2026.
• Reuters — Anthropic / Nscale, 26 Aug 2026.
• Reuters — Anthropic / Lambda, 31 Aug 2026.
• Reuters — CoreWeave, 11 Aug 2026.
• Microsoft — India South Central region, 6 Aug 2026.
• Microsoft — Saudi Arabia East, 31 Aug 2026.
• Reuters — IBM / Together AI, 11 Aug 2026.
• Cursor — joining SpaceX, 14 Aug 2026.
• OpenAI — decision on Cursor, 28 Aug 2026.
• Nasdaq — historical market data, retrieved 1 Sep 2026.

Research and Robotics

• Cruces et al. — education-based productivity gaps, submitted 4 Aug 2026.
• Yu et al. — workplace adoption and activity, submitted 16 Aug 2026.
• Davidovich et al. — Permission Denied, submitted 2 Aug 2026.
• Naik et al. — VAKRA, submitted 12 Aug 2026.
• PinSieve, submitted 27 Aug 2026.
• Reuters — Xpeng Robotics funding, 24 Aug 2026.
• Reuters — Tiangong Ultra, 25 Aug 2026.