MerlaTech | Stock Analysis: AI Signals

Micron released fiscal Q4 results on September 30 for a quarter ended September 3. For MU researchers, the useful question is how much reported revenue came from its two data center business units. These units are not a measure of AI revenue. The mix behind the $54.229B quarter Cloud Memory reported $16.283B and Core Data Center $18.002B. Together they contributed $34.285B, or 63.2% of $54.229B total revenue. In fiscal Q3, the same two units contributed $25.293B of $41.456B, or 61.0%. Their reported revenue share rose 2.2 percentage points. Figures and calculations use Micron’s segment table; shares are rounded. Fiscal quarter Cloud Memory + Core Data Center Total revenue Share Q3 FY26 $25.293B $41.456B 61.0% Q4 FY26 $34.285B $54.229B 63.2% Three checks for the next report Will the two-unit share hold or reverse in fiscal Q1? GAAP gross margin was 86.8% in Q4 versus 84.6% in Q3. Compare GAAP with GAAP. Q4 operating cash flow was $43.97B and net capex $10.77B. Micron’s $33.20B adjusted...

Meta’s September 29 launch matters to META investors because it moves Muse into a daily business workflow: a merchant can connect social accounts, a storefront and other tools, then ask an agent to analyze and draft work. The investor question is whether useful work becomes durable paid demand after the cost of delivering it. Meta has not reported that outcome in this announcement. Meta says Muse for Small Business can connect Instagram professional-account analytics, Facebook Pages and Meta ad accounts, as well as third-party services including Shopify, Intuit QuickBooks, Slack and Stripe. The company presents example tasks such as analyzing sales and campaigns or drafting next week’s content. It says nothing publishes, sends or spends without the user’s approval. These are announced product capabilities and examples, not our app tests or independently measured business results. The pricing distinction matters. Meta says Muse is free for most of what people need and offers subscriptio...

Why it matters: AI memory is getting faster, and testing it at production scale is a separate engineering challenge. Teradyne's September 29, 2026 Magnum E2 announcement describes an answer for advanced DRAM and flash interfaces. For TER investors, its supported protocols are only the first step: customer qualification, repeat production use and reported economics still need evidence. Original SignalSage research diagram. Source: Teradyne Investor Relations, Sep 29, 2026. This is a research illustration, not a Teradyne product image or an App test. Four interfaces in the launch The Teradyne investor release says Magnum E2 targets LPDDR6, DDR6, GDDR7 and NAND flash. Its Near-DUT architecture brings digital instruments close to the device under test. The company describes high parallelism and flexible NRZ, PAM3 and PAM4 signaling support. Engineering and production configurations share a hardware and software platform. These are company product descriptions; the release does not pr...

Trane Technologies announced two 250 MW AI factory cooling reference designs on September 28, 2026. For TT investors, the useful question is how a design claim becomes a customer order and then reported sales. The release describes engineering capacity; it does not say 500 MW of customer sites have been installed or sold. Original SignalSage comparison graphic. Tap for the full-size post. Metrics below are Trane's reference-design estimates, not measured field results. Two designs, two different resource questions Design #506 is air-cooled. Trane says its architecture could reallocate up to 22 MW of power to compute and improve annualized partial power usage effectiveness from 1.201 to 1.083 . Partial PUE covers the modeled portion of facility energy use; it is not a customer-wide performance result. Design #507 is modular. Trane says it could reallocate up to 8 MW and achieve design water usage effectiveness of zero using closed-loop dry coolers. That claim applies to the re...

Why this matters: NVIDIA announced an agent security platform on September 28, 2026. The launch addresses a real deployment question: who sets limits for an AI agent, and who notices when it crosses them? For NVDA investors, the separate question is whether that architecture eventually produces measurable revenue and margin. Source and timing: This note examines NVIDIA's September 28 investor release . The date is the announcement date. It is not an earnings report or a measured sales update. Two layers in the announcement OpenShell is open source runtime software that sets boundaries for agent actions, traces them, and enforces policy. NVIDIA says it runs with minimal overhead on Vera CPUs and can be extended to third-party compute platforms. Sentry is a reference system design for an independent watchdog on BlueField-4 DPUs. NVIDIA says it can detect an agent leaving its boundaries and quarantine it in milliseconds. Those are product descriptions and company claims, not evide...

Walmart’s CEO says AI shopping tools will not use a shopper’s income, history or urgency to raise an individual price. That is a meaningful trust promise, not a reported earnings result. In a September 25, 2026 letter , Walmart says Sparky can personalize help without personalizing price. The company commits not to vary prices by who a shopper is or the time of day, not to use information given to Sparky to raise prices or hide suitable lower-priced options, and to keep people overseeing pricing. These are company commitments, not an independent audit of transactions. What should a WMT investor watch? First, look for repeat visits and retention: a trusted shopping experience could support loyalty, but Walmart has not attributed those metrics to this statement. Second, watch digital baskets, order frequency and fulfillment economics if Sparky improves discovery. Third, compare sales growth with operating profit; better service can matter, but implementation costs matter too. None of the...

Original SignalSage by MerlaTech research diagram. Event announced Sept. 25, 2026. Amazon now offers a control room for sellers using its marketplace alongside eBay, Shopify, TikTok and Walmart. For AMZN investors, the important question is whether a simpler workflow leads to measurable Amazon service economics. Amazon's September 25, 2026 announcement says sellers can connect those four outside channels to Seller Central, link listings, view orders and profitability, and choose fulfillment. The rollout to U.S. sellers is gradual. Amazon says the new tools carry no additional cost to sellers. A product launch and a new connected merchant are therefore not, by themselves, reported incremental software revenue. Five storefronts, three separate ledgers Adoption: How many merchants connect channels, and how often do they actually use the tools? Amazon says more than 95% of independent sellers in its store already sell across multiple channels. That describes the potential workflow aud...

Original SignalSage research graphic. All example costs and acceptance rates are hypothetical. An agent that keeps working after you leave can also keep consuming resources. Microsoft’s September 25 Copilot announcement makes the billing question unusually concrete: everyday subscription work and long-running agent work have different spending models. Investors need to connect that architecture to customer value before jumping to a revenue conclusion. Microsoft says Cowork, Code and Autopilot run on usage-based billing. Autopilot is cloud-hosted and persistent; its private preview is expanding at month-end. Home and Code are beginning their Frontier rollout in the coming weeks. These are announced rollout stages, not proof that every customer can already use every feature. The useful unit is an accepted business outcome, rather than a task attempt. A supplier review, spreadsheet or internal tool might need rework, checking and approval. Counting every agent run as productive output hid...

Costco's membership fees were only about 2% of quarterly revenue, yet almost half the size of operating income. That comparison is useful. Calling it a membership-profit contribution is a different claim. In Costco's September 24 release, the 16-week fiscal Q4 ended August 30, 2026 had $1.850 billion in membership-fee revenue, $95.723 billion in total revenue and $3.801 billion in company operating income. Our calculations: 1.850 / 95.723 = 1.93%; 1.850 / 3.801 = 48.7%. Two denominators answer two different questions. The first is revenue mix. The second compares the size of one revenue line with company operating income. Neither measures the profit margin of a separate membership business. The release presents merchandise costs and selling, general and administrative expenses at company level, rather than a separate membership-profit calculation. A checkable operating-income bridge is $95.723B total revenue minus $83.531B merchandise costs minus $8.391B SG&A = $3.801B. ...

The next Micron EPS headline will be easier to judge if you prepare three questions first. Micron has scheduled its fiscal Q4 call for September 30, 2026 at 2:30 p.m. Mountain time. The calendar was announced August 26; the upcoming results are not yet reported in this note. 1. Revenue mix. Which business grew, and how much came from price, volume or product mix? Compare matching quarters and use the segment detail actually disclosed. Total memory sales do not reveal the contribution from every AI product. 2. Gross margin. Gross profit divided by revenue tells a different story from EPS. Check whether pricing and product mix cover manufacturing costs. Keep GAAP and adjusted figures separate and read the reconciliation. A margin change alone does not identify its cause. 3. Cash after capital spending. Read operating cash flow beside capital expenditures and working-capital changes. More capacity can consume cash before it generates sales. Use the company's stated free-cash-flow d...

Google's September 24 Demand Gen update reduces a specific step for some ads: a 9:16 full-bleed image ad on YouTube Shorts can send a viewer to the advertiser's landing page with one tap. That can make the path to a product page shorter. It does not, by itself, show that advertisers sold more products or that Alphabet earned more from ads. For an investor, there are two different ledgers. The advertiser pays for distribution and cares whether traffic turns into qualified actions. A useful test compares qualified conversions and cost per action on comparable campaigns, then checks whether the advertiser renews or expands its budget. A click is only an intermediate event. If it brings unqualified visits, the advertiser might spend less next quarter despite a higher click count. Alphabet's ledger is different. More convenient ad units could attract budget, but the proof would be incremental paid demand and durable ad revenue after considering inventory mix, pricing, delivery a...

The 407 ETR case is a clue, not revenue proof On September 22, 2026, Accenture Edge and AWS announced that six offers designed for mid-market organizations are generally available through AWS Marketplace. For investors, the question is whether buyers adopt them, renew them, and generate attractive profit after delivery costs. A Marketplace listing alone does not establish those outcomes. The six offers cover agent-led data discovery, AI-powered instance migration, customer-experience modernization, cloud-spend optimization, responsible-AI red teaming, and a secure cloud foundation. These are six routes into a buying conversation, not six disclosed revenue streams. The release includes one operating example. At 407 ETR, a contact-center modernization reportedly achieved 99.9% platform reliability and call and chat abandonment rates of 5% or less. That is evidence about this customer environment. It does not disclose contract value, Accenture revenue, AWS revenue, renewals, or contributi...

Qualcomm said on September 23 that it agreed to acquire PickNik, steward of the open-source MoveIt robotics framework. The transaction remains subject to closing conditions. The investor question is whether community reach can become shipped products, paid deployments and durable economics for Qualcomm. 1. The open community is a starting point Qualcomm says MoveIt 1 and MoveIt 2 will remain open source and continue to support a broad range of hardware. That openness could preserve adoption among researchers and robot makers. It also means an installed developer base is not, by itself, proprietary revenue. Watch for deal closure, ongoing releases, governance and evidence that contributors stay engaged. 2. Integrations must ship Qualcomm describes plans to connect the PickNik stack with its Dragonwing robotics portfolio and Arduino ecosystem. A roadmap differs from shipped tools. Useful follow-up evidence would show working releases, documentation, supported hardware, reliability in rea...

AWS made Claude Opus 5.5 available on Bedrock and Claude Platform on AWS on September 22. Anthropic estimates typical workloads cost 40% less to run than Opus 5. That is a customer-cost claim, not an AWS revenue or margin result. Who captures lower cost? Customers could run the same tasks and spend less. Lower costs could unlock more paid tasks. AWS benefits only if net billings and margins improve after model-provider and infrastructure costs. Availability alone cannot tell us which outcome wins. Research checklist Track paid task volume, net revenue per task, Bedrock adoption, retention and gross margin. Compare quality, latency, retry rates and cache use in real workloads. A repeat-workload cohort reveals whether lower cost stimulates demand or merely lowers the bill. Neither announcement quantifies the split among Anthropic, AWS and suppliers. The 40% figure is Anthropic's typical-run estimate versus Opus 5; results depend on prompt mix and caching. Neither announcement proves ...

NVIDIA released Isaac ROS 5.0 at ROSCon on September 22, 2026. Its open-source GPU-accelerated packages now add agent-ready development workflows and support ROS Lyrical and Ubuntu 24.04. NVIDIA says the release can run across Jetson devices. This is a developer milestone; the announcement does not report incremental robot sales or NVIDIA robotics revenue. Three rungs between a release and revenue The first rung is software availability: the packages, workflows and integrations must ship and be used. The second is deployment: developers need to convert pilots into paid robot designs, machines installed in the field and reliable operations. The third is economics: unit volumes, software and hardware attach rates, gross margins and repeat purchases determine whether a promising ecosystem becomes a durable business. Passing the first rung does not establish the third. NVIDIA says a new agent-ready FoundationPose inference library can estimate and track object pose up to 5.5 times fas...

Microsoft declared a quarterly dividend of $0.98 per share on 15 September 2026, up $0.07 from the prior quarter. That is about 7.69%, rounded to 8% in the company announcement. It is a change in the cash payment per share—not an 8% investment return. Three different numbers belong in your research note: • Quarterly cash: $0.98 per share. • Annualized run rate: $0.98 × 4 = $3.92, assuming four equal payments. This is not four future declarations or a guarantee. • Indicated yield: annualized dividend divided by the share price you actually use. Keep the price and its timestamp; this graphic assumes no live price. The announced ex-dividend and record date is 19 November 2026; the payment date is 10 December 2026. Those dates describe different events. A payment is not total return, which also depends on price changes and, for an individual, costs and taxes. Save the distinction before comparing dividend headlines. Start with a label before comparing a percentage A percentage needs a ...

AWS says Duke Energy cut data preparation tasks from two weeks to hours using its grid-planning AI agents. That is a specific workflow result—not a measured reduction in the time it takes to energize a project. The announcement is dated 17 September 2026. For a stock-research watchlist, separate three proof points: 1. Study ready: comparable cases, review time and reruns. 2. Build ready: approvals, equipment availability, crews and project economics. 3. Power on: actual service dates, usable capacity and full costs. A faster first step can help. It does not establish how much the whole project accelerates, or whether Amazon or a utility earns more. Engineers still make the final engineering decisions. Save this checklist for the next AI-infrastructure headline. Why the distinction matters An operational headline describes a unit of work. An investment thesis needs a chain of evidence. Before assigning a financial meaning to a time saving, identify exactly what started the clock, what s...

Adobe: $4.62 or $6.13 EPS? Read the Reconciliation Adobe reported Q3 FY2026 diluted EPS of $4.62 under GAAP and $6.13 on its non-GAAP basis. The quarter ended 28 August; the release was dated 10 September 2026. The per-share bridge is $4.62 + $1.38 stock-based/deferred compensation + $0.15 intangible amortization + $0.04 acquisition expenses − $0.05 investment gains/losses − $0.01 tax adjustments = $6.13. Figures are rounded as reported. The $1.51 gap is not extra cash paid to shareholders. Ask which excluded costs recur, how equity compensation affects dilution, what the cash-flow statement shows, and whether peer definitions match. Neither EPS measure alone establishes valuation or future returns. Source: Adobe SEC earnings release, 10 September 2026 SignalSage by the MerlaTech team — US stock research. English UI; some features require in-app purchases/subscriptions. Education, not investment advice. iPhone & Android: iPhone and Android download page Start with a definition befo...

TSMC reported August 2026 consolidated revenue of NT$514.806 billion in its 10 September release. July 2026 was NT$467.580 billion; August 2025 was NT$335.772 billion. These are reported revenue figures, not profit or stock returns. The month-over-month comparison adds NT$47.226 billion: 47.226 / 467.580 = about 10.1%. The year-over-year comparison adds NT$179.034 billion: 179.034 / 335.772 = about 53.3%. Both percentages describe the same August total, using different starting points. Do not add the two rates, subtract them to infer acceleration, or treat one month's growth as an annual forecast. Compare matching periods and currencies, then read margins, cash flow and the business mix separately. Save a three-field research note: reported period, comparison period, currency/unit. A growth percentage without its baseline is an incomplete fact. Start with the unit before doing any arithmetic. The release’s table uses NT$ millions. Dividing every amount by 1,000 produces the NT$...

Samsung + Mistral: when AI flags a chip defect, how often is it right? Samsung’s 9 September 2026 announcement describes plans to use on-premises AI in semiconductor operations, including defect detection and equipment optimization. It states goals, not a measured yield improvement. Here is our independent, fictional inspection run — NOT Samsung or Mistral performance data. Of 1,000 parts, 20 are defective and 980 are good. A detector catches 90% of defects (18) but also flags 5% of good parts (49). It produces 67 alerts, of which only 18, or 26.9%, correspond to real defects. Two defects are missed; 931 good parts are cleared. The denominator matters: detection rate is 18/20; alert precision is 18/67. Neither alone establishes manufacturing value. A flagged part is not automatically scrapped: review, retesting and process decisions change the economic result. Save three questions for the next factory-AI headline: What is the real defect base rate? How much review and downtime do ...

NVIDIA’s +24% token-throughput case: where is the revenue gate? NVIDIA’s 15 September 2026 post describes Lambda’s DSX MaxLPS test using 19 HGX B200 nodes versus a 16-node baseline within the same power budget. It reports 24% more cluster token throughput. That is a specific company-reported test, not a universal result or a revenue increase. Our separate fictional acceptance test: A generates 1,000,000 tokens and 80% pass a fixed acceptance rule, leaving 800,000 accepted tokens. B generates 1,240,000 and 60% pass that same rule, leaving 744,000. Raw output is 24% higher, while accepted output is 7% lower. These invented acceptance rates are NOT Lambda or NVIDIA measurements. They illustrate why a throughput number alone is insufficient; they do not allege quality deterioration in the reported test. Save three gates before turning an efficiency headline into a stock thesis: 1. Accepted work: same model, quality threshold, latency target and system boundary? 2. Paid demand: how much ...

Microsoft’s AI sales case: 20% higher close rates—than what? Microsoft’s 17 September 2026 post reports a sales-team comparison. Its footnote describes internal data on 687 sellers from January–June 2024, comparing regular Copilot users with low-usage sellers. This is a company-reported case, not today’s sales result or an isolated causal estimate of AI’s effect. Our fictional ruler starts at a 10% close rate. A 20% relative increase reaches 12%: a gain of 2 percentage points. Adding 20 percentage points instead reaches 30%, a 200% relative increase. These are teaching examples, NOT Microsoft’s actual rates. Before turning a productivity headline into a stock thesis, save three questions: What is the baseline, population and measurement period? What else differed between the groups? Did additional wins become recognized revenue with attractive margins after software, training and review costs? A useful operational signal still needs a bridge to business economics. Source: https://blogs...

Amazon announced on 10 September 2026 that it was adding six Ariane 6 launch contracts, expanding its Arianespace commitment from 18 to 24 launches. That is a useful fact about contracted capacity. It is not evidence that those six additional missions have already flown, or a measure of subscriber revenue. The announcement is a starting point for a research notebook. To understand the business implications, separate four milestones that are often compressed into a single growth headline: booked, deployed, service ready, and paid usage. 1. Booked: a commitment with conditions A signed launch commitment can secure access to a scarce operational resource. The next questions concern execution: what delivery windows apply, what flexibility exists, and what dependencies could change the schedule? A headline contract count does not disclose all payment terms, cancellation provisions or the allocation of risk between counterparties. Record what the source actually says in one column. Put the u...

A manufacturing milestone needs a denominator. Intel and ASML’s 7 September 2026 company update says High-NA EUV had processed more than one million wafers across early certification and testing, R&D, and select-layer volume production for a subset of Panther Lake products. This article examines the scope of that statement; it does not independently verify the companies’ results. Schematic research graphic. The wafer grid does not represent the share of testing, R&D or production. What the total can tell you The statement describes accumulated processing activity spanning several purposes. It gives a useful starting point for asking where a technology has been used. Its categories matter: certification work, development work and production work answer different questions. Combining them into one count does not reveal the contribution of each category. Before comparing this figure with another announcement, record the measurement period, system boundary and counting method. Ask...

By the SignalSage / MerlaTech team. Apple set 18 September 2026 as the first-market availability date for iPhone 18 Pro and Pro Max. Its 9 September announcement and 12 September preorder date establish a launch calendar, not reported sales. This research note asks what a sold-out label could establish, and what remains unknown. We are not reporting observed Apple stockouts. Two stores, one sign, different scale Imagine Store A receives 20 phones and 30 purchase requests. Store B receives 200 phones and 210 requests. Assume each request is for one phone, there is no replenishment during the observation period, and requests are distinct and valid. Both stores exhaust their inventory. Both also have 10 requests beyond available stock. The fulfilled quantities differ tenfold: 20 versus 200. Requests divided by starting inventory are 1.5 and 1.05. Those ratios describe this invented setup; neither is a company growth rate or a margin estimate. A binary sold-out label discards most of this ...

A quantum milestone deserves a resource budget beside it. Before turning a laboratory result into a business conclusion, write down the task, the accuracy and what the complete workflow consumed. IBM’s research update dated 15 September 2026 describes a continuum between error mitigation and error correction. Its argument centers on useful computation rather than a logical-qubit count alone. Our graphic is a conceptual reading guide, not measured results. Three resource questions Hardware: How many physical qubits and supporting operations does the demonstrated circuit require? Keep the useful task separate from the machinery needed to protect it. Repetition and runtime: How many runs are needed to estimate the answer at the specified accuracy? A short individual run is not automatically a short complete job. Classical compute and control: What processing happens around the quantum circuit? Record what is included in the reported time and what is excluded. An incomplete boundary m...

By the SignalSage / MerlaTech team. Which AI chip wins? A headline number cannot answer that until you know what the test measured. MLCommons released MLPerf Inference 6.1 on 16 September 2026, including end-to-end RAG and edge agentic tests. These workloads make the comparison question more specific, not less. Watch the ranking change One fictional example, two different questions Assume the same fictional workload and quality target. System A completes 800 jobs per second using eight devices. System B completes 1,200 using sixteen. B has 50% more total throughput. Divide by device count: A = 800 / 8 = 100 jobs per second per device. B = 1,200 / 16 = 75. A leads on this normalized measure, even though B leads on total output. These numbers are invented for explanation, not MLPerf measurements or a comparison of actual vendors. Neither ranking establishes a cheaper system. Device prices, host equipment, software, electricity, networking and utilization remain unknown. Nor does either r...