AI, Big Tech & Hyperscaler Radar · September 2026
The Cost of Making AI Work
Cheaper model access enlarges the opportunity. Financing and verified execution determine how much becomes durable value.
1–30 September 2026 · Research completed 1 October · Full monthly editionDr. Michael Schymura
Read the evidence50–55% lowerSol / Luna successor token baskets; no quality adjustment
76 / 100Agent Reality Index; up from 72, editorial judgment
56 source recordsFinancials, releases, policy, research and market data
CASEFILE
Three tests guide the evidence: financing, accepted-task cost and verification. Reported results, derived sensitivities and future plans are distinguished.
01The constraint is moving beyond the model
September made advanced AI cheaper to buy and harder to assess as a business system. New models reduced the price of useful inference. Supplier results demonstrated substantial demand. Yet the same month made the dependence on customer financing, controlled access and verification more visible.
Oracle supplies the clearest accounting example. Its reported quarterly operating cash flow was $23.103bn. Of that, $11.363bn came from customer prepayments with a significant financing component. Cash capital expenditure was $28.499bn. The resulting reported free cash flow was −$5.396bn; subtracting those advances as a sensitivity produces −$16.759bn. This is a way to expose the financing mix, not a correction to Oracle’s accounting. S01 S02
The strategic question therefore changes. A lower token bill can improve the economics of a task. It cannot, by itself, establish that the task completes reliably, that customers will keep paying, or that infrastructure commitments can be funded through a difficult demand cycle. Those are separate tests, and they belong in separate exhibits.
This edition’s central judgment is that verification is becoming a productive input. Permissions, review, evaluation and cash-flow discipline enable deployment. They also consume resources. The firms that measure these costs alongside model prices will make better allocation decisions than those that treat capability announcements as realized returns.
Scope: events and material publications from 1–30 September 2026, Europe/Berlin. Research was completed on 1 October. Later retrieval is distinguished from the observation window; the legacy token-price basket is explicitly provisional.
Cheaper model access enlarges the opportunity. Financing and verified execution determine how much becomes durable value.
02Ten developments that changed the assessment
Signal scores rank strategic materiality, not publicity or investment attractiveness. A score of 1 denotes noise or a weak signal; 3 a meaningful sector development; 5 a development capable of changing system economics. Availability, evidence quality and time horizon remain separate.
| Rank | Development | Signal | Maturity at cutoff | Evidence |
|---|---|---|---|---|
| 1 | Oracle cash-flow bridge | 5/5 | reported financials | S01 S02 |
| 2 | Cheaper GPT-6 inference | 4/5 | released; priced | S06 S07 |
| 3 | Memory and custom-silicon revenue | 5/5 | recognized supplier revenue | S03 S04 |
| 4 | Astra and controlled frontier access | 4/5 | staged / restricted rollout | S05 S09 |
| 5 | Agent runtimes reach GA; workflows remain mixed | 4/5 | GA runtime; workflow previews | S12 S13 S14 |
| 6 | Patent drafting experiment | 4/5 | working-paper RCT | S18 |
| 7 | China’s integrated roadmap | 4/5 | announced / in training | S36 |
| 8 | European financing and industrial ambition | 4/5 | financed / proposed | S19 S20 S24 |
| 9 | Meta extends its distribution strategy | 3/5 | roadmap / organization | S37 S38 |
| 10 | Verification becomes infrastructure | 4/5 | released controls / prospective assurance | S27 S32 S34 S35 |
The ordering gives recognized financial outcomes more weight than roadmaps. It also gives a bounded randomized experiment more weight than a larger collection of demonstrations. The detailed ledger in radar.json records confidence, horizon and an invalidation condition for every entry. Scores are editorial judgments; a new source can alter them without changing the underlying event.
03Three structural changes, three common overclaims
First, falling inference prices and rising infrastructure commitments can coexist. They describe different margins: the cost of another request and the obligation to finance capacity. There is no contradiction to resolve by placing both on a decorative dual-axis chart.
Second, the agent market is developing an operating layer. Runtime isolation, identities, persistent context, approvals and audit trails are appearing across competing platforms. These are necessary complements to model capability. Their presence is evidence of product maturation, though it is not evidence that every workflow has become autonomous or economical. S13 S14 S27
Third, distribution is widening. Enterprise suites, developer platforms, browsers and wearable devices offer different ways to reach users. The important question is which channel produces repeated, paid, controllable use. An installed audience is an option on adoption, not a conversion rate. S12 S21 S37 S38
Three claims deserve resistance. A larger model under training is not a released product. A multi-gigawatt target is not energized capacity. A percentage improvement on a vendor’s task set is not a percentage improvement in the purchaser’s business. Each claim can be useful once its denominator, timing and maturity are preserved.
04Competition reaches into the operating environment
Microsoft’s advantage remains its position inside work: documents, communication, identity and established enterprise distribution. September’s Copilot announcements point toward more persistent delegation, but the rollout language matters. Home and Code were entering Frontier deployment; Autopilot was expanding private preview. Procurement should test the features actually available to its tenant. S12
AWS concentrated on the environment in which agents run. Runtime V2 reached general availability in five regions; Bedrock Managed Agents, jointly developed with OpenAI, entered a three-US-region preview. These are different readiness levels. Google’s Gemini 3.8 Live strengthened the real-time voice interface, where latency and conversational continuity matter alongside reasoning. S13 S14 S15
Oracle’s revenue evidence is stronger than a roadmap, while its funding structure deserves its own examination. Meta’s enterprise initiative seeks to turn its models, agents and business relationships into a broader software business. Its consumer reach offers a distribution route, but enterprise permissioning, support and retention still have to be demonstrated. S01 S38
Alibaba’s strategy spans models, chips, context services and the agent operating environment. This makes the competitive unit increasingly the complete system. IBM and Huawei remain in the scorecard, with August ratings carried forward where this run did not establish a comparable September change. An unchanged rating records the evidence boundary; it is not a claim that those companies were inactive.
05Customer advances matter to the funding bridge
The Oracle bridge distinguishes three things that are easily conflated: cash generated in the reported operating category, cash invested in property and equipment, and the timing support provided by customers. Advances can be commercially valuable. They can align demand with buildout and reduce near-term external financing needs. They can also concentrate exposure to a small group of counterparties. S02
The useful stress test is transparent: hold every other reported flow fixed and remove the $11.363bn advance contribution. This produces the sensitivity shown in Exhibit 1. It does not predict the next quarter and does not assume that advances vanish. It asks how much the observed cash-flow position depends on this source of finance.
Reuters’ report of $518bn in Anthropic commitments over a decade adds a second, less directly inspectable layer. The underlying private filing was unavailable for this edition, so the amount remains attributed reporting. It is not added to August’s industry lease total. A customer commitment, a supplier receivable and a data-centre lease can describe related claims along the same chain. Summing them would create a misleading impression of independent investment. S39
For boards, the required reconciliation is practical: committed capacity, financing secured, energized capacity, utilized capacity, recognized revenue and collected cash. A business can improve on one line while becoming more exposed on another.
Inspect the cash-flow bridge
Customer advances are part of the reported result. Compare that result with a sensitivity that removes them while holding every other cash flow fixed.
Cash-flow view
Pinned reported result: −$5.396bn. Sensitivity excluding $11.363bn customer advances: −$16.759bn. Neither view forecasts a future quarter.
Exhibit 01S02
Customer advances change the financing picture
USD billions · Oracle FY27 Q1
Inspect exact values and units
| Component | USD bn |
|---|---|
| Operating cash flow | 23.103 |
| of which customer advances | 11.363 |
| Cash capex | 28.499 |
| Reported free cash flow | −5.396 |
| Sensitivity excluding advances | −16.759 |
Why this form. The components reconcile exactly; sequence preserves the accounting identity. Challenger considered. sankey diagram (rejected): Sankey lacks observed counterparty flows; reject inferred topology.
06Capability advances arrive with access conditions
Astra’s September rollout extended the frontier while retaining staged access and safeguards for sensitive capabilities. Fable and Mythos 5.1 similarly illustrate that access policy can separate products built on closely related underlying capability. These are economically relevant product characteristics, since a purchaser buys a usable service under particular terms. S05 S09
Later releases compressed the distance between high-end capability and broadly priced inference. GPT-6.1 Sol followed GPT-6 Sol and Luna within the month. Anthropic released Opus 5.5 and Sonnet 5.5. Google’s live models address another margin: interactive performance in a voice session. S06 S07 S10 S11 S15
This radar does not combine their published benchmarks into a league table. Tool access, time budgets, harnesses, context length and task composition differ. A benchmark can establish improvement within a carefully matched evaluation; it does not automatically support a universal ranking across products.
The next useful comparison is a frozen set of the buyer’s own tasks, with the same acceptance criteria, escalation rules and latency limits. Record both failures and successful outputs. Restricting the sample to completed tasks would hide precisely the reliability cost the comparison is intended to reveal.
07The price of tokens is only one part of the task bill
At an 80% input / 20% output mix, the listed GPT-6 Sol basket costs $3.60 per million total tokens, compared with $7.20 for the August GPT-5.6 Sol price. The analogous Luna comparison is $0.18 against $0.40, a 55% reduction under this basket. These are successor comparisons, with no claim of identical quality or token requirements. S06 S44
GPT-6.1 Sol preserves the $2 input / $10 output standard tariff while reducing cached-input pricing to $0.10. Opus 5.5 lists at $4/$20; Sonnet 5.5 at $2/$10. Anthropic’s reported Sonnet cost-per-task improvement is a token-efficiency claim, not a reduction in those standard list prices. Cache hits, batching and discounts require separate scenarios. S07 S10 S11
August’s fixed Terra/Luna cohort remains at 50 on its original launch-price base of 100. This is a provisional carry-forward: documents accessed on 1 October match the earlier prices, but this run did not recover an independently archived September 30 snapshot. Substituting the new Luna into the old series would manufacture a like-for-like price decline. S42 S43 S44
The buyer’s denominator should be an accepted task. Add model charges, tools and retrieval, retries, human review, monitoring and support. A cheaper model is economical when the entire workflow remains cheaper at the required quality. The package preserves the token arithmetic while leaving that broader outcome to deployment evidence.
Exhibit 02S06 / S44
Successor token baskets are 50–55% cheaper
USD per million total tokens · 80% input / 20% output
Legacy fixed-cohort index: 50, provisional carry-forward; October 1 current documents corroborate unchanged legacy list prices.
Inspect exact values and units
| Family | August basket | September basket | Change |
|---|---|---|---|
| Sol | $7.20 | $3.60 | −50% |
| Luna | $0.40 | $0.18 | −55% |
Why this form. Two points and an explicit connector expose the before/after comparison. Challenger considered. radial lollipop chart (rejected): Radial length and angle reduce price precision without adding evidence.
08Availability is becoming more concrete than autonomy
The runtime layer made a verifiable advance. AgentCore Runtime V2 is generally available; its provider reports lower cold-start latency across tested package sizes. Cold starts are one component of an agent run, so the result should not be described as an equivalent reduction in end-to-end workflow time. S13
The application layer remains mixed. Managed-agent previews, Copilot rollout stages and DevDay’s expanding agent interfaces create useful options, but they do not have identical regional coverage, service guarantees or administrative controls. The readiness matrix records these differences instead of counting each launch as a production deployment. S08 S12 S14
There are three procurement questions. Can the agent act only within an explicitly defined authority? Can an operator inspect, interrupt and recover its work? Can the owner measure accepted output against the full cost of operating it? A demonstration that answers only the first happy-path task leaves the investment question open.
The Agent Reality Index rises from 72 to 76 under August’s weights. That is a modest editorial revision, driven by runtime availability, governance and bounded outcome evidence. Its reliability component remains just 34 out of 100. The aggregate should not conceal the weakest part of the system.
09Supplier earnings locate the current scarcity rents
Broadcom’s AI semiconductor revenue reached $16.7bn in its reported fiscal quarter, within total revenue of $29.591bn. Micron reported $54.229bn in quarterly revenue. Their cash-flow profiles differ sharply from Oracle’s because their business models and investment responsibilities differ. Exhibit 3 keeps the reported fiscal periods visible. S03 S04
These results support strong demand for parts of the compute stack. They do not reveal the returns earned by the businesses buying the infrastructure or the welfare gains enjoyed by users. Revenue recognized by a supplier is a cost somewhere else in the chain. The distribution of profit can therefore remain highly uneven even while usage grows.
Alibaba’s V900 announcement adds a future alternative, with commercial release planned for Q1 2027. Its performance claims and large-cluster architecture remain vendor statements until representative deployments can be examined. Qwen 4 was in training; later models with much larger parameter counts were roadmaps. S36
AMD’s agreement to acquire World Labs brings model research closer to hardware and system design. That may improve insight into emerging workloads. The proposed transaction is also an expensive strategic option whose payoff depends on integration and future products, rather than an immediately measurable increase in accelerator revenue. S29 S30
Exhibit 03S02 / S03 / S04
The cash burden sits in different places
USD billions · reported fiscal quarters differ
Oracle FY27 Q1
Broadcom FY26 Q3
Micron FY26 Q4
Inspect exact values and units
| Company | Revenue | OCF | Capex | FCF | Period |
|---|---|---|---|---|---|
| Oracle | 19.345 | 23.103 | 28.499 | −5.396 | FY27 Q1 |
| Broadcom | 29.591 | 14.197 | 0.532 | 13.665 | FY26 Q3 |
| Micron | 54.229 | 43.970 | 10.770 | 33.200 | FY26 Q4 |
Why this form. Separate equal-scale panels preserve company identity and fiscal-period caveats. Challenger considered. sankey diagram (rejected): Intercompany cash flows are not observed, so a flow diagram would fabricate links.
10The grid constraint has an operational margin
The IEA’s September grid report supplies a useful counterweight to the assumption that every capacity problem requires a new power plant and a new transmission line. It estimates that grid-enhancing technologies could unlock up to 330GW of additional connection capacity globally. The estimate covers generation, storage and demand together. It is not a pool reserved for AI facilities. S26
The implication is institutional as well as technical. Dynamic ratings, better visibility and power-flow management may improve the use of existing assets, but utilities must procure, operate and trust these systems. A technically available option does not become deliverable capacity until the connection and operating arrangements exist.
For an AI project, the meaningful milestone is therefore not the announcement of a site or an electricity agreement alone. It is a sequence: connection terms, firm or flexible supply, network work, equipment delivery, commissioning and usable compute. Each milestone has a different failure mode and financing requirement.
September’s European and Chinese capacity targets belong on a forward watchlist. They cannot be used to lower a current infrastructure-pressure score merely because their announced size is large. The index remains near its ceiling, while the IEA evidence identifies a route by which pressure could eventually ease. S20 S36
11Usage is spreading, with a large measured gap
Microsoft’s September diffusion publication estimates that 18.8% of the world’s working-age population used AI in June. Its Global North and Global South estimates are 28.8% and 16.2%, respectively: a gap of 12.6 percentage points. These are adjusted telemetry-based estimates of use, released later than the observation period. S16
That distinction limits the conclusion. The figures show uneven reach under the report’s measurement framework. They do not show a September acceleration, regular use by every counted individual, the intensity of use, paid conversion or productivity. Changes in tool coverage can also affect comparability between editions.
Distribution partnerships can help reduce some adoption friction. The Mozilla–Mistral beta is one example of AI reaching users through an existing application; its German rollout was still planned for later in the year. Meta’s device strategy pursues a different route through persistent consumer access. S21 S37
The practical adoption funnel should remain disaggregated: eligible users, active users, repeated use, accepted tasks and retained paying accounts. A system can gain exposure while failing at the later stages. That is why diffusion and commercialization are separate parts of this radar.
Exhibit 04S16
A 12.6-point North–South usage gap
Share of working-age population, % · June 2026 estimates
Inspect exact values and units
| Region | Share, % |
|---|---|
| Global South | 16.2 |
| World | 18.8 |
| Global North | 28.8 |
Why this form. Common positions compare three percentages without implying causal ordering. Challenger considered. waffle chart (rejected): A waffle adds counting effort and area without improving the comparison.
12A bounded experiment is more informative than an unbounded promise
The September patent-drafting working paper offers unusually useful evidence: 133 lawyers at eleven US firms were randomized over three months. Expert-rated drafting improved by 0.34 standard deviations at day 10 and 0.38 at day 90. The corresponding outcome samples were 105 and 98. Exhibit 5 reconstructs approximate 95% intervals from rounded, individual-clustered standard errors. S18
The paper also examines unassisted redlining and reports a positive average effect, with heterogeneous estimates by experience. However, compliance with the no-AI instruction could not be directly enforced. Attrition, task scope and the short follow-up limit generalization. The study does not establish an economy-wide employment effect or permanent preservation of expertise. S18
Microsoft’s internal planning-cycle case points in the same direction of possible workflow improvement but has a weaker identification design: selected before/after observations without random assignment. It is useful for formulating an operating hypothesis, not for estimating an average treatment effect across firms. S17
The management response should preserve both kinds of evidence. Use cases locate promising work; controlled comparisons test whether the gains survive a counterfactual. Track quality and learning as well as speed. A team that completes drafts faster but becomes less able to judge them may have shifted costs into a later stage.
Exhibit 05S18
Drafting quality gains survive three months
Control-group SD units · expert ratings; approximate 95% intervals
Inspect exact values and units
| Task | Effect (SD) | Approx. 95% CI | n |
|---|---|---|---|
| Drafting: day 10 | +0.34 | [0.03, 0.65] | 105 |
| Drafting: day 90 | +0.38 | [0.09, 0.67] | 98 |
Why this form. Point estimates with intervals are the evidence; baseline includes zero. Challenger considered. violin plot (rejected): Participant-level distributions are unavailable; a violin would invent density.
13Capital buys options; operating results determine their value
Mistral’s announced €3bn Series D, at a post-money valuation above €21bn, gives the European ecosystem greater financial room to build and distribute models. The funding amount and the valuation describe different quantities. They should not be added together, converted into an implied revenue figure or treated as equivalent to a compute commitment. S19
AMD’s approximately $8.2bn World Labs transaction is all-stock and remained subject to closing conditions. The agreement was signed on September 26 and announced on September 28. The acquisition broadens AMD’s research capabilities; the economic outcome depends on what that research changes in future systems and products. S29 S30
The reported Anthropic obligations illustrate the other side of capital access: long-lived promises can accumulate before the associated revenue is earned. Contract duration, cancellation rights, minimum consumption and counterparty concentration matter more than a single headline total. S39
This edition does not publish a global September AI-funding total. The available announcements differ in currency, instrument, stage and disclosure completeness. The transaction ledger provides identifiable events with their status; an incomplete but inspectable set is more useful than a precise-looking aggregate without a defensible coverage rule.
14Implementation advances without a new-law count
The AI Board’s September meeting addressed enforcement and implementation, including frontier incidents and transparency. Its institutional significance lies in coordination and the development of supervisory practice. The meeting record is not an enforcement sanction. S23
The Commission’s copyright consultation is a request for evidence and views, with a November 3 response deadline. It is relevant to training data, licensing and downstream content businesses, but it does not itself change the law. The IPCEI AI process concerns the design and pre-notification of potential support, with assessment still ahead. S24 S25
These distinctions explain why the Regulatory Pressure Index remains at August’s 86. Existing implementation demands continue, but this run did not establish a comparable September change in binding obligations, fines or export-control coverage that warranted a new score.
For deployment teams, the useful work is concrete: identify provider and deployer roles, document the system’s intended use, maintain traceable versions and assign an owner to supervisory requests. The radar reports policy developments; it does not substitute a general announcement for a jurisdiction- and use-specific legal assessment.
15The disclosure date is not the incident date
Anthropic’s September report describes selected operations observed from December 2025 through August 2026. OpenAI’s September 30 distillation disclosure describes a July campaign. Both are September information events. Neither is evidence that all described incidents occurred during September. S34 S35
The reports also lack a population denominator. More detected cases can reflect more use, better detection, more capable attackers or changed disclosure policy. Their strategic value lies in the failure modes and defensive lessons, rather than a fabricated monthly incident-growth rate. Attribution should remain attached to the organization making the assessment.
NVIDIA’s Open Agent Safety announcement addresses runtime containment, testing and monitoring. It is evidence that infrastructure vendors recognize the problem. It is not a warranty that a control will hold under every adversarial workflow or third-party integration. S27
The deployment test should include authority boundaries, data access, untrusted tool output, interruption and recovery. Evaluate the whole operating environment, including partner-hosted services. A strong model safeguard can coexist with an unsafe permission configuration. Conversely, a bounded and observable workflow can support useful deployment before general reliability is solved.
16Control options improve without eliminating dependence
Open models and open operating tools provide different kinds of flexibility. Model weights can support local inference or alternative hosting. A runtime can improve containment and observability. Neither automatically grants the training data, complete reproducibility, unrestricted commercial rights or the ability to operate cheaply at scale.
Mistral’s financing and distribution partnerships strengthen a European option. NVIDIA’s OpenShell and Isaac ROS work expand parts of the software environment. Alibaba’s full-stack roadmap adds competition but contains a mixture of released capabilities and future commitments. Their licensing and availability have to be evaluated at the specific component level. S19 S21 S27 S28 S36
The Open-Source Pressure Index remains 90 after rounding. A small improvement in enterprise adoption evidence does not justify inventing a new download or benchmark series. Those components are explicitly carried forward from August until comparable evidence becomes available.
For buyers, an option has value when it can be exercised. Test exportable data, portable evaluation sets, reproducible configurations and a credible alternative operator. Merely having access to weights may leave dependence concentrated in accelerators, networking, specialist staff or a proprietary application layer.
17Discovery still needs characterization
Anthropic’s enzyme-system announcement is a useful early example of model-assisted scientific search. Its own account leaves a central question unresolved: the full function of the identified system. That boundary should remain attached to the result. A novel pattern, a laboratory observation, a biological mechanism and a therapeutic application are different evidentiary stages. S31
The Life Sciences Verification Program addresses who may access certain capabilities, while the Accenture arrangement seeks to expand evaluation capacity. Both suggest that research productivity and assurance are developing together. Neither establishes the accuracy of a scientific result by itself. S32 S33
The economic opportunity is to reduce the cost of proposing, screening and refining hypotheses while preserving the independence of the experimental test. If model suggestions expand the number of plausible candidates faster than laboratories can validate them, the bottleneck may move into instrumentation, experimental design or expert interpretation.
October’s evidence threshold is therefore replication and characterization, not another promotional description. The radar will update the science assessment when the chain from computational suggestion to independently supported mechanism becomes clearer.
18The integration layer is advancing faster than the return evidence
Isaac ROS 5.0 provides a concrete software event for robotics: new integration capabilities and an expanded development environment. AMD’s proposed World Labs acquisition reflects the strategic value assigned to spatial models and simulation. Mistral’s Munich initiative links AI development to industrial and physical applications. S20 S28 S29
These developments support an inference about where firms expect value to emerge. They do not establish that general-purpose robots can be deployed economically across factories. Performance in a bounded manipulation task leaves maintenance, safety, changeovers, exception handling and uptime unresolved.
A robotics business case needs a stable unit of output and the full operating cost of achieving it. Measure intervention frequency, useful operating hours, recovery time and the engineering effort required to transfer the system to another site. A demonstration video cannot supply those denominators.
No new sector-wide installed-fleet or independently audited robotics-return series was verified in this run. The section therefore records software and strategic milestones while keeping scaled operational productivity as an open test.
19Equities discriminate across the AI complex
The selected public-equity basket did not move as one trade. AMD and Meta led the September price gains, while several cloud and semiconductor names declined. This is consistent with investors differentiating among expectations and exposures; it is not a causal estimate of the effect of any launch or earnings announcement.
| Security | Aug 31 close, USD | Sep 30 close, USD | Price return | Versus SPY, pp |
|---|---|---|---|---|
| AMD | 470.72 | 611.76 | +29.96% | +30.54 |
| META | 572.34 | 725.18 | +26.70% | +27.28 |
| MU | 958.73 | 1065.11 | +11.10% | +11.67 |
| NVDA | 220.78 | 228.38 | +3.44% | +4.02 |
| GOOGL | 339.35 | 344.08 | +1.39% | +1.97 |
| MSFT | 507.29 | 512.90 | +1.11% | +1.68 |
| SPY | 767.05 | 762.63 | -0.58% | +0.00 |
| AMZN | 259.77 | 249.15 | -4.09% | -3.51 |
| AVGO | 370.34 | 351.19 | -5.17% | -4.59 |
| BABA | 114.02 | 107.54 | -5.68% | -5.11 |
| IBM | 233.87 | 219.93 | -5.96% | -5.38 |
| ORCL | 149.12 | 137.30 | -7.93% | -7.35 |
Source: the two regular-close observations in each Stock Analysis history table, supplied by S&P Global Market Intelligence S45 S46 S47 S48 S49 S50 S51 S52 S53 S54 S55 S56. Returns exclude dividends and after-hours trading. The Close column is used rather than Adjusted Close; the provider may adjust historical prices for splits. This preserves August’s price-return definition but changes the data vendor from Nasdaq. No splice is used to manufacture an August–September return.
Micron’s September 30 results came after the regular close, so that day’s closing price cannot be read as a reaction to those results. Broader interest-rate, macroeconomic and company-specific news also affects returns. Huawei is unlisted; model laboratories without listed shares do not receive synthetic equity returns.
Exhibit 06S45 / S46 / S47 / S48 / S49 / S50 / S51 / S52 / S53 / S54 / S55 / S56
The AI equity basket did not move together
Price return, % · August 31 to September 30 regular close
Inspect exact values and units
| Security | Aug 31 close | Sep 30 close | Return, % | Versus SPY, pp |
|---|---|---|---|---|
| AMD | 470.72 | 611.76 | +30.0 | +30.5 |
| META | 572.34 | 725.18 | +26.7 | +27.3 |
| MU | 958.73 | 1065.11 | +11.1 | +11.7 |
| NVDA | 220.78 | 228.38 | +3.4 | +4.0 |
| GOOGL | 339.35 | 344.08 | +1.4 | +2.0 |
| MSFT | 507.29 | 512.90 | +1.1 | +1.7 |
| SPY | 767.05 | 762.63 | −0.6 | 0.0 |
| AMZN | 259.77 | 249.15 | −4.1 | −3.5 |
| AVGO | 370.34 | 351.19 | −5.2 | −4.6 |
| BABA | 114.02 | 107.54 | −5.7 | −5.1 |
| IBM | 233.87 | 219.93 | −6.0 | −5.4 |
| ORCL | 149.12 | 137.30 | −7.9 | −7.4 |
Why this form. Zero-based signed lengths reveal positive and negative returns. Challenger considered. radial lollipop chart (rejected): Area makes exact returns harder to compare and lacks a second observed dimension.
20Optimism is becoming more conditional
The editorial mood assessment moves from +2 in August to +1 on the same −5 to +5 scale. The selected reporting remains constructive about AI demand, while financing and the gap between growth and expectations receive greater attention.
Three headlines illustrate the tension: Reuters reported Broadcom guidance below estimates, Micron guidance above estimates, and Anthropic’s large, substantially binding buildout commitments. This is a deliberately small, disclosed sample. It is not a representative media corpus, an automated sentiment index or a measure of public opinion. S39 S40 S41
The recurrent narrative error is to treat every positive quantity as the same form of validation. Funding validates investors’ willingness to provide capital. A contract validates a commitment. A shipment validates delivery. Recognized revenue validates an accounting event. Retention and cash collection answer further questions.
The mood score is useful only as a compact description of this reading of the month. It should not be correlated with stock returns as if it were a statistically constructed independent series. A future corpus-based measure would require a documented methodological break.
21Europe gains options before it gains independence
Mistral’s financing and German expansion, the IPCEI AI design and Telekom’s industrial agenda give Europe a more substantial set of options. Their stages differ: new capital, a regional operating presence, a policy process and an infrastructure expansion agenda. None is interchangeable with a completed sovereign technology stack. S19 S20 S22 S24
The German case is particularly useful because industrial demand can be specific. Engineering, production, maintenance and scientific workflows offer tasks with measurable outputs. They also demand integration with existing systems and strong operational responsibility. That makes local expertise and customer access valuable even when accelerators or some software layers are imported.
The European Dependency Index stays at 84 after rounding. A modest improvement in the foundation-model component does not outweigh persistent reliance on accelerators, large cloud environments and developer platforms. The 1GW German compute ambition is dated to 2030; it is not credited as September capacity.
The practical sovereignty test is the ability to continue operating, change suppliers and control sensitive work under realistic constraints. Locate the data, identify the operator, inspect the contract, test export and restore procedures, and measure the switching cost. A regional label is the beginning of that inquiry.
22Read the verb before counting the launch
Exhibit 7 records a small cross-section of September developments using their actual maturity. A generally available runtime, a US-only preview, an announced acquisition and a capacity target are all material; they answer different questions.
| Claim | What is supported | What remains open |
|---|---|---|
| AgentCore Runtime V2 | GA in five regions | End-to-end workflow performance |
| Bedrock Managed Agents | Three-region US preview | Broad availability and durable economics |
| Copilot Autopilot | Expanded private preview | Production coverage and measured outcomes |
| Mistral German compute | 1GW ambition for 2030 | Financing, connection and delivered capacity |
| Alibaba V900 | Q1 2027 release plan | Shipment, system performance and adoption |
| Muse on glasses | Integration planned over coming months | Availability and repeated use |
| AMD–World Labs | Signed acquisition agreement | Approval, closing and integration |
The general rule is to retain the narrowest verb supported by the source. “Plans” should not become “builds”; “available” should not become “adopted”; “adopted” should not become “profitable.” Subsequent editions can move an item through those stages when evidence arrives. S12 S13 S14 S20 S29 S36 S37
Exhibit 07S12 / S13 / S14 / S20 / S29 / S36 / S37
The readiness verbs are different
Selected September announcements · status at cutoff
Inspect exact values and units
| Product | Status at cutoff | Date | Source |
|---|---|---|---|
| AgentCore Runtime V2 | GA in five regions | 2026-09-18 | S13 |
| Bedrock Managed Agents | US preview | 2026-09-29 | S14 |
| Copilot Autopilot | Private preview | 2026-09-25 | S12 |
| Mistral German compute | 1GW target for 2030 | 2026-09-28 | S20 |
| Alibaba V900 | Q1 2027 planned release | 2026-09-22 | S36 |
| Meta Muse on glasses | Coming months | 2026-09-23 | S37 |
| AMD–World Labs | Signed; closing pending | 2026-09-28 | S29 |
Why this form. A text table retains nominal maturity and future dates without inventing a numerical scale. Challenger considered. alluvial diagram (rejected): No longitudinal cohort transitions exist for an alluvial flow.
23Eight indices, with August’s weights preserved
These 0–100 indices are structured editorial judgments, not measured probabilities, physical quantities or statistically validated composites. Components and weights are included in the data file. Higher is not uniformly better: infrastructure pressure, regulatory pressure and European dependency represent constraints or exposure.
| Index | July | August | September | Monthly change |
|---|---|---|---|---|
| Innovation | 82 | 89 | 92 | +3 |
| Commercialization | 77 | 85 | 86 | +1 |
| Infrastructure Pressure | 89 | 94 | 95 | +1 |
| Regulatory Pressure | 77 | 86 | 86 | +0 |
| Market Expectations | 72 | 90 | 91 | +1 |
| Open-Source Pressure | 80 | 90 | 90 | +0 |
| Agent Reality | 60 | 72 | 76 | +4 |
| European Dependency | 86 | 84 | 84 | +0 |
Innovation and agent reality improve, but the latter remains constrained by reliability. Infrastructure pressure edges higher. Commercialization gains are supported primarily by recognized supplier revenue and bounded cases, with retention carried forward. Regulatory pressure and European dependency do not change at displayed precision.
Small movements should not be overinterpreted. An illustrative five-point change in a component weighted at 20% moves its index by one point. That sensitivity is often as large as the displayed monthly change. The methodology therefore records unchanged components and their carry-forward status, rather than treating every component as freshly measured.
Exhibit 08S44 / S01 / S05 / S13 / S19
Innovation and agent readiness move most
Editorial indices, 0–100 · August to September 2026
Inspect exact values and units
| Index | July | August | September | Change |
|---|---|---|---|---|
| Innovation Index | 82 | 89 | 92 | +3 |
| Commercialization Index | 77 | 85 | 86 | +1 |
| Infrastructure Pressure Index | 89 | 94 | 95 | +1 |
| Regulatory Pressure Index | 77 | 86 | 86 | 0 |
| Market Expectations Index | 72 | 90 | 91 | +1 |
| Open-Source Pressure Index | 80 | 90 | 90 | 0 |
| Agent Reality Index | 60 | 72 | 76 | +4 |
| European Dependency Index | 86 | 84 | 84 | 0 |
Why this form. Common 0–100 positions preserve direction and size of judgment changes. Challenger considered. radar chart (rejected): Polygon area blends indices with different normative directions.
24Capability breadth does not settle the return question
The scorecard retains the sixteen August dimensions, including model capability, cloud infrastructure, proprietary silicon, enterprise distribution, security, governance and sovereign-cloud position. Each is rated from 1 to 10; the mean is an equal-weight summary of ordinal assessments.
| Company | Mean rating, 1–10 | September assessment |
|---|---|---|
| Microsoft | 8.69 | score unchanged |
| Amazon | 8.81 | score unchanged |
| 8.81 | score unchanged | |
| Meta | 6.56 | score unchanged |
| Oracle | 8.06 | modest strengthening |
| IBM | 7.94 | score unchanged |
| Alibaba | 8.31 | score unchanged |
| Huawei | 7.62 | score unchanged |
Oracle’s cloud-infrastructure component rises from 8 to 9 on revenue and delivery evidence. Other components remain unchanged because September information did not justify a full one-point revision. IBM and Huawei have explicit carry-forward assessments. The exact means are recalculated from the original component values, rather than from August’s rounded averages.
Exhibit 9 shows every component so that the mean does not hide strategic differences. Meta’s distribution and open-model position coexist with a different enterprise and cloud footprint. A high integrated score is not an estimate of valuation, expected return or business quality. The scorecard describes capabilities; the cash-flow evidence tests a different dimension.
Exhibit 09S44 / S01 / S12 / S13 / S15 / S36 / S38
Capabilities differ across sixteen dimensions
Editorial ratings, 1–10 · September; full labels in exact-value table
| Company | Mean | 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 | sovereign cloud position |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Microsoft | 8.69 | 9 | 9 | 9 | 7 | 9 | 9 | 10 | 10 | 9 | 9 | 9 | 9 | 6 | 7 | 10 | 8 |
| Amazon | 8.81 | 8 | 9 | 10 | 9 | 9 | 9 | 9 | 7 | 9 | 9 | 9 | 9 | 8 | 9 | 10 | 8 |
| 8.81 | 9 | 9 | 9 | 10 | 9 | 9 | 8 | 9 | 9 | 9 | 8 | 9 | 8 | 9 | 10 | 7 | |
| Meta | 6.56 | 8 | 8 | 2 | 8 | 5 | 9 | 7 | 4 | 6 | 5 | 5 | 8 | 10 | 8 | 8 | 4 |
| Oracle | 8.06 | 7 | 8 | 9 | 7 | 9 | 7 | 9 | 8 | 8 | 8 | 10 | 9 | 6 | 8 | 8 | 8 |
| IBM | 7.94 | 7 | 7 | 8 | 6 | 8 | 7 | 9 | 7 | 9 | 10 | 9 | 7 | 9 | 7 | 8 | 9 |
| Alibaba | 8.31 | 9 | 10 | 9 | 10 | 8 | 8 | 8 | 7 | 7 | 7 | 8 | 8 | 10 | 10 | 7 | 7 |
| Huawei | 7.62 | 7 | 7 | 8 | 9 | 7 | 7 | 8 | 7 | 8 | 8 | 9 | 7 | 8 | 8 | 5 | 9 |
Inspect exact values and units
| Company | Mean | 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 | sovereign cloud position |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Microsoft | 8.69 | 9 | 9 | 9 | 7 | 9 | 9 | 10 | 10 | 9 | 9 | 9 | 9 | 6 | 7 | 10 | 8 |
| Amazon | 8.81 | 8 | 9 | 10 | 9 | 9 | 9 | 9 | 7 | 9 | 9 | 9 | 9 | 8 | 9 | 10 | 8 |
| 8.81 | 9 | 9 | 9 | 10 | 9 | 9 | 8 | 9 | 9 | 9 | 8 | 9 | 8 | 9 | 10 | 7 | |
| Meta | 6.56 | 8 | 8 | 2 | 8 | 5 | 9 | 7 | 4 | 6 | 5 | 5 | 8 | 10 | 8 | 8 | 4 |
| Oracle | 8.06 | 7 | 8 | 9 | 7 | 9 | 7 | 9 | 8 | 8 | 8 | 10 | 9 | 6 | 8 | 8 | 8 |
| IBM | 7.94 | 7 | 7 | 8 | 6 | 8 | 7 | 9 | 7 | 9 | 10 | 9 | 7 | 9 | 7 | 8 | 9 |
| Alibaba | 8.31 | 9 | 10 | 9 | 10 | 8 | 8 | 8 | 7 | 7 | 7 | 8 | 8 | 10 | 10 | 7 | 7 |
| Huawei | 7.62 | 7 | 7 | 8 | 9 | 7 | 7 | 8 | 7 | 8 | 8 | 9 | 7 | 8 | 8 | 5 | 9 |
Why this form. Printed values retain precision across 128 ordinal cells; tint is only secondary. Challenger considered. radar chart (rejected): Eight radial polygons obscure individual ratings and invite area ranking.
25Four decisions for the next operating cycle
For the CFO, reconcile AI commitments with financing and utilization. The first deliverable is a table that connects contract duration, cancellation rights, customer advances, cash capex and collected revenue. Stress the timing of cash receipts and commissioning separately. A useful stop condition is a project whose financing assumes utilization that no signed customer schedule supports.
For the CIO, move two or three valuable workflows through a common evaluation. Freeze acceptance criteria, measure total cost per accepted task and retain a human escalation route. Use the released feature set in the actual deployment region. Expand only after the quality and operating-cost thresholds hold outside the demonstration sample.
For the security owner, test permissions, interruption, logging and recovery as part of the product. Assign an accountable operator to the system’s actions. A failed containment test should narrow authority before a broader rollout, rather than being averaged away inside a general model score.
For European industrial leaders, use the new options to test portability. Select a workflow where data control, local integration or continuity has economic value, and compare at least one alternative operating arrangement. The relevant output is a measured switching path and a credible operating plan. Capacity announcements and funding rounds are inputs to that decision, not its conclusion.
26October tests and the August watchlist
The next edition should focus on evidence that resolves a specific uncertainty.
| October test | Evidence that would change the assessment |
|---|---|
| Customer-financed infrastructure | Conversion of advances into service revenue, delivery and utilization; concentration disclosures |
| Astra and successor models | Independent evaluations with matched harnesses, budgets and acceptance criteria |
| Managed-agent previews | Dated GA, regional coverage, service terms and named production outcomes |
| Enterprise workflow economics | Accepted output, retries, human review and retained usage across repeated cycles |
| European compute projects | Connection, financing and commissioning milestones beyond capacity targets |
| World Labs acquisition | Regulatory progress or closing; integration plans tied to concrete products |
| Research productivity | Replication, attrition analysis and longer follow-up in expert tasks |
| Copyright consultation | Submissions before November 3 and subsequent official conclusions |
August’s Astra watch is partly resolved: a staged September release occurred, with constraints still material. The Anthropic financing watch gained attributed reporting but not a publicly inspectable prospectus. The EU enforcement watch produced implementation coordination, not a verified fine in this run. Independent evaluation remains open. Power-flexibility production evidence remains incomplete. Cursor’s November 12 watch remains a future checkpoint. The Nscale/Lambda and other reported-deal watches are unresolved here; no closing is inferred. S05 S23 S26 S39 S44
The most consequential weak signal is the convergence of product design and assurance. If verification becomes cheaper and more reproducible, more tasks become economic at a given model quality. If verification costs rise as delegation deepens, token deflation may produce much less savings than the price table suggests.
27Evidence, limitations and reproducibility
The package contains 56 source records, 49 screened candidate records, a claim ledger, original analytical data, executable calculations and nine paired figures. Primary financial releases, official product documentation, policy records and research are prioritized. Reported private documents remain distinctly labeled.
The month is defined by Europe/Berlin calendar boundaries. A September publication about an earlier observation is included as a new information event, with its observation period preserved. Announced future capacity and future availability remain prospective. The only post-window price check is the October 1 corroboration of the legacy cohort, explicitly identified as provisional.
August’s original JSON and Markdown were recovered unchanged. They provide numerical continuity, but the supplied archive did not contain the full prior source ledger, chart assets or QA evidence. July values are inherited from August’s recorded comparison, not independently reconstructed. Six prior full editions were not available; no missing history is fabricated.
Run code/reproduce.py to recalculate cash-flow identities, token baskets, market returns and indices from data/radar-data.json. The accompanying build sources create the figures, offline article and editable carousel. Source URLs and retrieval references are recorded in sources.csv; chart tables preserve exact values. Raw copyrighted articles are not redistributed.
Source register
- S01
Oracle
Oracle Q1 FY2027 earnings2026-09-10financial resultprimary-verified
Observation: FY27 Q1 ended 2026-08-31. Boundary: Reported quarter; backlog is neither revenue nor cash.
- S02
Oracle
Oracle Q1 financial tables2026-09-10financial tablesprimary-verified
Observation: FY27 Q1 ended 2026-08-31. Boundary: Prepayment exclusion is a sensitivity, not a GAAP correction.
- S03
Micron
Micron fiscal Q4 and FY2026 results2026-09-30financial resultprimary-verified
Observation: FQ4 ended 2026-09-03; 14 weeks. Boundary: Fiscal quarter differs from Oracle and Broadcom; not AI application profitability.
- S04
Broadcom
Broadcom fiscal Q3 results2026-09-02financial resultprimary-verified
Observation: FY2026 Q3. Boundary: Company segment classification; not a measure of final-user AI surplus.
- S05
OpenAI
GPT-6 Astra2026-09-03technical releaseprimary-verified
Observation: September launch. Boundary: Vendor evaluations; eligibility and product rollout are separate from capability.
- S06
OpenAI
GPT-6 Sol and Luna2026-09-22price and releaseprimary-verified
Observation: September list prices. Boundary: Uncached standard API prices; workload quality, retries and review excluded.
- S07
OpenAI
GPT-6.1 Sol2026-09-29price and releaseprimary-verified
Observation: September list prices. Boundary: Different release channels; no independent blanket quality ranking.
- S08
OpenAI
DevDay 2026 recap2026-09-29product releaseprimary-verified
Observation: September event. Boundary: Count availability by feature and customer class, not by event headline.
- S09
Anthropic
Claude Fable and Mythos 5.12026-09-01technical releaseprimary-verified
Observation: September launch. Boundary: Some enterprise safeguards were scheduled for later in the fall.
- S10
Anthropic
Claude Opus 5.52026-09-22price and releaseprimary-verified
Observation: September list prices. Boundary: Vendor workload savings are distinct from list-price reductions.
- S11
Anthropic
Claude Sonnet 5.52026-09-28price and releaseprimary-verified
Observation: September list prices. Boundary: Vendor task mix; not a 30% universal price reduction.
- S12
Microsoft
New Copilot with Home, Code and Autopilot2026-09-25product roadmapprimary-verified
Observation: September announcement. Boundary: An announcement does not establish universal September general availability.
- S13
AWS
AgentCore Runtime V2 generally available2026-09-18product releaseprimary-verified
Observation: September GA in five regions. Boundary: Vendor cold-start tests do not measure complete agent-task latency.
- S14
AWS
Bedrock Managed Agents preview2026-09-29product previewprimary-verified
Observation: September preview; three US regions. Boundary: US preview availability is not EU general availability.
- S15
Google
Gemini 3.8 Live2026-09-15technical releaseprimary-verified
Observation: Published September 15; updated September 17. Boundary: Vendor benchmarks and product-specific availability; not independent field performance.
- S16
Microsoft
Continued state of global AI diffusion2026-09-21measurement reportprimary-verified
Observation: June/Q2 2026 usage; released September. Boundary: Adjusted telemetry, not a random survey or September productivity estimate.
- S17
Microsoft
Microsoft Customer Zero lessons2026-09-17internal case studyprimary-verified
Observation: Five planning cycles April–August 2026. Boundary: Selected internal before/after case; no randomized counterfactual.
- S18
Autor et al. / MIT
Does AI assistance enhance or erode expertise?2026-09working paperprimary-verified
Observation: Three-month randomized study; September paper. Boundary: Outcome sample 105/98; AI use in unassisted redlining could not be directly enforced.
- S19
Mistral AI
Mistral Series D2026-09-08funding announcementprimary-verified
Observation: September financing. Boundary: Funding is a flow; valuation is a stock; neither is revenue.
- S20
Mistral AI
Hallo Deutschland2026-09-28investment roadmapprimary-verified
Observation: September announcement; 2030 capacity target. Boundary: Target, not operating capacity; location does not eliminate equipment dependence.
- S21
Mistral AI
Mistral and Mozilla2026-09-16product partnershipprimary-verified
Observation: September beta; later-2026 German rollout. Boundary: German availability remains a future milestone.
- S22
Deutsche Telekom
Digital X 20262026-09-09corporate announcementprimary-verified
Observation: September event. Boundary: Existing German AI infrastructure is not a new September opening.
- S23
European Commission
Ninth AI Board meeting2026-09-18official meeting recordprimary-verified
Observation: Meeting September 17; published September 18. Boundary: Meeting record, not a fine or a newly enacted obligation.
- S24
European Commission
First IPCEI on AI design2026-09-16official policy processprimary-verified
Observation: Design September 16; updated September 17. Boundary: Pre-notification is not state-aid approval or completed infrastructure.
- S25
European Commission
Technology and copyright consultation2026-09-29official consultationprimary-verified
Observation: September 29 opening; November 3 closing. Boundary: Consultation, not a legal change or compliance deadline.
- S26
IEA
Modernising Grids in the Age of Electricity2026-09-21analytical reportprimary-verified
Observation: September report; modelled potential. Boundary: Modelled technical potential across all uses, not allocated AI data-centre capacity.
- S27
NVIDIA
Open Agent Safety platform2026-09-28technical releaseprimary-verified
Observation: September announcement and OpenShell availability. Boundary: Reference designs and vendor tests are not certified universal containment.
- S28
NVIDIA
Isaac ROS 5.02026-09-22technical releaseprimary-verified
Observation: September ROSCon release. Boundary: Software release and task demonstrations do not establish fleet reliability or factory ROI.
- S29
AMD
World Labs acquisition agreement2026-09-28M&A announcementprimary-verified
Observation: Agreement September 26; announced September 28. Boundary: Signed, not closed; stock consideration is not a cash funding round.
- S30
AMD / SEC
World Labs agreement Form 8-K2026-09-28regulatory filingprimary-verified
Observation: Agreement September 26. Boundary: Share count depends on the contractual pricing mechanism near closing.
- S31
Anthropic
Claude enzyme-system research2026-09-23research announcementprimary-verified
Observation: September early research. Boundary: Early laboratory work and preprint-stage evidence, not a clinical outcome.
- S32
Anthropic
Accenture embedded evaluation2026-09-18assurance partnershipprimary-verified
Observation: Five-year prospective commitments. Boundary: Prospective capacity; Anthropic funding creates an independence question to manage.
- S33
Anthropic
Life Sciences Verification Program2026-09-17controlled-access betaprimary-verified
Observation: September beta. Boundary: Access screening is a governance mechanism, not proof of safety or research success.
- S34
Anthropic
September threat intelligence report2026-09-10vendor incident disclosureprimary-verified
Observation: Observed December 2025–August 2026. Boundary: Selected detected cases; no population denominator or September incident rate.
- S35
OpenAI
Coordinated model-distillation campaign2026-09-30vendor incident disclosureprimary-verified
Observation: July activity; September disclosure. Boundary: Attribution is the vendor assessment; attempts are not all successful extractions.
- S36
Alibaba Cloud
Alibaba full-stack AI roadmap2026-09-22technical roadmapprimary-verified
Observation: September event; 2027/2032 targets. Boundary: Roadmap and vendor performance claims; future scale is not installed capacity.
- S37
Meta
Meta Connect 2026 recap2026-09-24product roadmapprimary-verified
Observation: Connect September 23; recap September 24. Boundary: Integration roadmap, not universal September availability.
- S38
Meta
Meta Enterprise Platform2026-09-28organizational announcementprimary-verified
Observation: September launch of enterprise initiative. Boundary: Organization and distribution strategy; no disclosed enterprise revenue outcome.
- S39
Reuters
Anthropic buildout commitments2026-09-29reported private documentsreported
Observation: Reported decade commitments. Boundary: Underlying private filing unavailable; reported, not independently audited here; no aggregation with counterparties.
- S40
Reuters
Broadcom outlook and market expectations2026-09-02market reportingreported
Observation: September 2 report. Boundary: Selected headline; not a representative sentiment sample.
- S41
Reuters
Micron outlook2026-09-30market reportingreported
Observation: September 30 report. Boundary: Selected headline; results announced after the regular-market close.
- S42
OpenAI
GPT-5.6 Terra pricing documentationundatedcurrent price documentationcarry-forward
Observation: Retrieved October 1; prior-cohort corroboration. Boundary: No independently archived September 30 snapshot; carry-forward provisional.
- S43
OpenAI
GPT-5.6 Luna pricing documentationundatedcurrent price documentationcarry-forward
Observation: Retrieved October 1; prior-cohort corroboration. Boundary: No independently archived September 30 snapshot; carry-forward provisional.
- S44
Michael Schymura
August radar edition2026-09-26prior editioncarry-forward
Observation: August observation window; package recovered. Boundary: Prior ZIP contained Markdown and JSON only; no complete prior figure/QA archive.
- S45
Stock Analysis / S&P Global Market Intelligence
SPY historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S46
Stock Analysis / S&P Global Market Intelligence
MSFT historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S47
Stock Analysis / S&P Global Market Intelligence
NVDA historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S48
Stock Analysis / S&P Global Market Intelligence
AMZN historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S49
Stock Analysis / S&P Global Market Intelligence
GOOGL historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S50
Stock Analysis / S&P Global Market Intelligence
META historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S51
Stock Analysis / S&P Global Market Intelligence
ORCL historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S52
Stock Analysis / S&P Global Market Intelligence
IBM historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S53
Stock Analysis / S&P Global Market Intelligence
BABA historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S54
Stock Analysis / S&P Global Market Intelligence
AVGO historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S55
Stock Analysis / S&P Global Market Intelligence
AMD historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.
- S56
Stock Analysis / S&P Global Market Intelligence
MU historical prices2026-09-30market-data tableobserved
Observation: Regular closes August 31 and September 30 2026. Boundary: Close field, not dividend-adjusted total return; vendor changes from August Nasdaq source.