If you’re in NYC today, come find us at the Data Management Summit. Edge & Node CEO @rodventures takes the stage for a keynote at 3:05 PM, followed by our very own Craig Tutterow joining a panel at 3:25 PM. We also have a booth in The Gallery, we’d love to meet you. Secure your delegate pass here: #DMSNYC# #DataManagementSummitNYC# #DataManagement# #DataStrategy# #CapitalMarketsTech# #InstitutionalTrading#

The Data Management Summit NYC is tomorrow 🗽 3:05 @rodventures keynotes on more data at lower cost, and the case for AI-enabled specialization. 3:25 Craig Tutterow joins the panel on building a defensible audit trail for agentic workflows. @ Ease, 605 Third Ave, 7th floor. Come find us 🤝 #DMSNYC#

Our CEO @rodventures is keynoting the Data Management Summit New York on Sept 17. More Data, Lower Costs: The Case for AI-Enabled Specialization. AI is making more data. AI is also making the systems that handle it cheaper to build. He'll show how those two things meet at scale. 3:05 PM. NYC. Hope to see you there

💡With SSI (Self-Sovereign Identity) data management technology, individuals, as the subjects of their information, can provide only the data they choose, enabling self-directed management of personal information.🛡️

The Edge & Node team will be at the Data Management Summit NYC this Thursday. 3:05 : @rodventures keynotes on more data at lower cost, and the case for AI-enabled specialization. 3:25 : Craig Tutterow joins the panel on building a defensible audit trail for agentic workflows. @Ease, 605 Third Ave. #DMSNYC# We hope to see you there!

Craig Tutterow, our Director of Data, takes the stage at 3:25 PM on September 17th at the Data Management Summit New York. The panel: Proving Your AI: Building a Defensible Audit Trail for Agentic Workflows in a Regulated Environment. Regulators are asking firms to show their work on AI decisions. Most governance frameworks were built for a different era. Craig and the panel get into what a defensible audit trail looks like now. 40 minutes, moderated by Pete Harris of the A-Team Group.

# Practical and Useful Patterns with ADK State manages your agent's "current situation." With dynamic prompts and scoped data management, build truly intelligent conversations 🎯 📌 **Title**: State 🔗 **URL**: ## 🧩 Overview State is the working memory accessible via `session.state`. It holds "data needed right now in this conversation" -- like shopping cart contents or information gathered step by step during a booking flow. State's key feature is **prefix-based scoping**: - `app:` - Data shared across the entire app (global settings for all users) - `user:` - Data shared per user (user preferences, persisted across sessions) - `temp:` - Temporary data valid only within the current session - No prefix - Standard session scope Additionally, embedding `{state_var}` in agent instructions enables dynamic prompts that adapt in real time. ## 🛠 How to Use Set values on `session.state` using dictionary-style access -- for example, `session.state["booking_date"] = "2026-07-15"`, `session.state["party_size"] = 4`, and `session.state["seat_type"] = "window"` for standard session-scoped data. For scoped state, use the `user:` prefix (e.g., `"user:preferred_language"`) for per-user persistent data, `temp:` (e.g., `"temp:search_results"`) for data that disappears when the session ends, and `app:` (e.g., `"app:maintenance_mode"`) for data shared across all users. Additionally, embed placeholders like `{booking_date}`, `{party_size}`, and `{user:preferred_language}` in the `Agent`'s `instruction` parameter to dynamically inject state values into the agent's prompt at runtime. ## 🏗 Practical Usage **Restaurant booking flow:** Gather information progressively through conversation, accumulating in State: 1. "I'd like a reservation for 4 tomorrow evening" → Save `booking_date`, `party_size` to State 2. "Window seat please" → Add `seat_type` 3. "I'm allergic to nuts" → Add `allergies` 4. At final confirmation, pull all info from State to generate a confirmation message With `{booking_date}` embedded in instructions, the agent always responds with awareness of the latest booking status. **Persisting user preferences:** Using the `user:` prefix, carry forward information like "this person prefers English" or "this person is vegetarian" even when sessions change. ## 💡 Use Cases - 🛒 E-commerce: Manage cart contents in State. `temp:` for search filters, `user:` for shipping address - ✈️ Travel booking: Progressively accumulate dates, party size, seat type in State - 🤖 Personalization: Learn and persist user preferences via `user:` scope - 📊 Dashboards: Share system-wide status via `app:` scope ## ⚠️ Caveats - Storing large amounts of data in State degrades performance. Consider Artifacts for bulky data - Changes to `app:` scope affect all users. Use with caution - When embedding `{state_var}` in instructions, unset values may display the placeholder as-is. Provide defaults - Establish key naming conventions to avoid collisions ✨ Leverage State's scoped management and dynamic prompts to build context-aware, intelligent agents. Perfect for booking flows and personalization! #ADK# #AIAgent#

Can dense, sparse and multi-vector approximate nearest neighbor search live in a single Rust codebase? Modular and Efficient Approximate Nearest Neighbors Search with Vectorium and kANNolo 🧩 Overview Two Rust libraries: Vectorium for data management and kANNolo for indexing. Encoders and datasets are compile-time type parameters, so each combination is specialized at build time. 🎯 The problem ・Faiss has a separate code path per index/quantizer combination, which is hard to extend ・Other libraries factor out only the graph or only the quantization ・None handles all three embedding types behind one interface 🛠 Method ・11 encoders (PQ, RaBitQ, sparse, multi-vector, and more) ・Hnsw and IVF, plus a two-stage RerankIndex that pairs any first-stage index with any rescoring ・A structure-of-arrays PQ table layout makes table building 4.5-5.2x faster 📊 Results ・Up to 3.26x hnswlib's throughput at 99% accuracy (kANNolo-Rerank, dense data) ・Smallest memory on every sparse collection ・But Glass wins on AgNews, Seismic is faster on sparse data, and builds are not fast #ANNSearch# #Rust#

$FFIV KEY READ-THROUGHS FROM F5’S Q4 FY2026 EARNINGS CALL F5’s October 27 call supports a selective broadening of enterprise AI spending beyond GPUs into data delivery, runtime security, traffic governance, and sovereign infrastructure. The strongest positive read-throughs concern memory pricing, enterprise storage, independent AI-security platforms, and infrastructure that makes shared AI capacity commercially usable. The principal negative implication is that strong infrastructure demand can coexist with weaker cash conversion and limited earnings leverage as suppliers accumulate inventory and absorb component inflation. The distinction between market growth and company-specific capture is critical: F5’s competitive wins do not identify the displaced public companies, its AI bookings are not recurring revenue, and its sovereign deployments do not establish incremental GPU orders. Impact assessments below distinguish business-level significance from consolidated earnings materiality. Country labels identify corporate domicile. SEMICONDUCTORS AND ENTERPRISE HARDWARE MEMORY PRICING HAS SUPPORT OUTSIDE HYPERSCALER AI, BUT INVENTORY PURCHASES ARE NOT END-CONSUMPTION (READ-THROUGH 1) Affected companies: Micron Technology (MU: US) and Sandisk (SNDK: US). Direction: positive for pricing durability and supplier bargaining power. Magnitude: moderate as corroborating industry evidence; negligible direct earnings significance from F5 alone. The supporting call evidence is unusually concrete. F5 attributed its FY2027 gross-margin outlook of 80.5% to 82.5%, versus 84.0% in FY2026, to higher component costs. Management identified “memory and storage” as the principal supply concerns and described continued inventory accumulation, supplier prepayments, and procurement for next-generation appliances. Inventory increases are expected to be strongest in the first half of FY2027. The transmission mechanism extends beyond the familiar demand for memory inside GPU servers. Enterprise networking and application-delivery appliances also consume memory and storage components. When these buyers secure supply earlier, prepay suppliers, and hold larger inventories, component manufacturers gain better order visibility and potentially stronger pricing. Micron has exposure through DRAM and NAND; Sandisk through NAND-based storage. F5 did not identify its suppliers, so neither company should be assigned a specific F5 contract or revenue contribution. The non-consensus implication is that a slowdown in an individual equipment replacement cycle need not immediately weaken upstream component orders. F5 expects replacement bookings to decline in the second half of FY2027, yet is still increasing inventory and purchasing components for future appliances. Supplier purchases can therefore remain stronger than contemporaneous equipment demand because procurement includes supply insurance and next-generation preparation. There are important limits. Advance procurement shifts demand between periods and can eventually create an inventory correction; it does not establish equivalent growth in underlying memory-bit consumption. This is also not an HBM-specific signal. Moreover, F5’s FY2027 gross-margin range improved by 50 basis points versus its earlier framework. The call confirms persistent cost pressure, not a newly worsening shock of the full year-over-year magnitude. Near term, Sandisk’s October 29 results provide the clearest trading catalyst for testing pricing, bookings, and customer procurement behavior. Micron’s September 30 results had already established a strong memory-cycle baseline, making F5’s disclosure corroboration rather than discovery. The longer-duration implication is continued component-supplier pricing support into the first half of 2027, with customer inventory accumulation becoming an increasingly important indicator of eventual cycle risk. SERVER AND INFRASTRUCTURE REVENUE GROWTH CAN OVERSTATE CASH EARNINGS QUALITY (READ-THROUGH 2) Affected companies: Dell Technologies (DELL: US) and Hewlett Packard Enterprise (HPE: US). Direction: negative for cash-conversion assumptions and the valuation premium assigned to revenue growth without equivalent cash generation. Magnitude: moderate, particularly in hardware-intensive businesses; insufficient evidence for an automatic EPS reduction. F5 reported Q4 operating cash flow of $118 million and free cash flow of $91 million against adjusted net income of approximately $255 million. Management attributed the cash weakness to inventory and supplier prepayments. It also maintained an annual pricing-review cadence despite rapidly changing component costs. These disclosures illustrate how strong demand can require additional cash before the associated revenue is collected. The transmission mechanism to Dell and HPE is shared exposure to constrained components and large hardware fulfillment obligations, not a supplier relationship with F5. Higher component prices increase the cash required to support a given physical shipment volume. Longer procurement lead times require earlier commitments. If customer pricing, supplier payments, and customer collections adjust at different speeds, revenue and accounting profit can increase while free cash flow underperforms. Cost pass-through creates a separate analytical issue. Higher equipment selling prices can lift revenue without a comparable increase in physical deployments or economic value added. Even full dollar recovery of component inflation does not necessarily preserve the gross-margin percentage. Conversely, favorable inventory costs and rapid repricing can temporarily improve margins. Revenue growth, unit growth, gross profit dollars, and working-capital investment therefore need to be evaluated separately. This is not evidence that Dell and HPE are already experiencing the same outcome. Dell’s September release showed very strong revenue and earnings growth, while HPE’s latest reported quarter generated approximately $1.0 billion of free cash flow, increasing approximately $0.2 billion year over year. HPE’s cash performance is explicit counterevidence against mechanically transferring F5’s deterioration. F5’s annual bonus-payment change is also company-specific and should not be imposed on either peer. Near-term catalysts are the next results covering their fiscal quarters ending in October 2026, particularly inventory, purchase commitments, pricing realization, backlog conversion, and cash-flow guidance. The longer-duration implication is a wider valuation distinction between infrastructure suppliers that translate demand into cash and those that finance growth through progressively larger balance sheets. The high-conviction negative read-through concerns earnings quality and capital intensity, not a forecast of weakening infrastructure demand. ENTERPRISE STORAGE AND DATA INFRASTRUCTURE INFERENCE DEPLOYMENT SUPPORTS ENTERPRISE DATA INFRASTRUCTURE, NOT JUST GPU PROCUREMENT (READ-THROUGH 3) Affected company: NetApp (NTAP: US). Direction: positive. Magnitude: moderate for enterprise AI and hybrid-cloud demand; modest incremental information relative to NetApp’s already strong reported performance. F5 disclosed more than $100 million of direct AI bookings for FY2026, increasing nearly 300%, and stated, “The majority of that is actually AI data delivery.” Customers deployed 29.2 terabits of F5-powered AI data-delivery capacity in Q4, increasing nearly 250%. Management also highlighted S3 capabilities, storage partnerships, and demand led by banking, financial services, and healthcare. The mechanism is the requirement to make enterprise datasets securely accessible to inference infrastructure. GPU availability alone does not resolve storage throughput, application access, availability, or policy constraints. Production deployments can require upgrades to object storage, data management, and the infrastructure connecting storage to compute. The opportunity is therefore broader than adding storage capacity: it includes improving the performance and governability of existing enterprise data. NetApp’s exposure is supported by an identified relationship rather than thematic similarity. Its established collaboration with F5 combines NetApp data infrastructure, including StorageGRID, with F5 traffic management and security for S3-based AI workflows. F5’s stronger AI data-delivery activity consequently validates a use case that NetApp already addresses commercially. The geographic evidence strengthens the durability argument. F5 reported 31% EMEA growth and described customers repatriating applications to local clouds or private data centers because of sovereignty requirements. That supports infrastructure capable of operating across private and public environments. It is not evidence that public-cloud storage must contract: NetApp’s own latest release showed growth in both Hybrid Cloud and Public Cloud revenue. The magnitude must remain bounded. F5’s deployed throughput is not a measure of NetApp shipments, storage capacity purchased, or exclusive supplier share. S3 is an ecosystem standard, and F5 supports multiple storage vendors. Some expenditure can improve access to existing storage rather than create a new array purchase. Near term, NetApp’s next FY2027 update provides a test of whether AI-related pipeline activity converts into additional bookings and revenue. Its September 2 release had already reported 30% revenue growth, so F5’s call supports persistence rather than predicting an unreported quarter. The longer-duration positive is enterprise inference spending becoming less narrowly concentrated in accelerators and more dependent on governed data infrastructure. AI NETWORKING AND ACCELERATED INFRASTRUCTURE BLUEFIELD GAINS ARCHITECTURAL VALIDATION AS AI FACTORIES BECOME MULTITENANT SERVICES (READ-THROUGH 4) Affected company: NVIDIA (NVDA: US). Direction: positive for BlueField’s role in enterprise and sovereign AI infrastructure. Magnitude: meaningful product-level validation, but immaterial near-term consolidated earnings impact from the disclosed projects. The direct supporting evidence is Telenor’s selection of F5 BIG-IP Next for Kubernetes running on NVIDIA BlueField-3 DPUs. The call described tenant isolation, policy-based traffic control, and observability operating in the AI service path while preserving host CPU capacity. This is an explicitly identified technology relationship, not an inferred supplier allocation. The transmission mechanism is the transition from dedicated GPU capacity to shared commercial services. Bare-metal GPU rental requires less application-level governance than an infrastructure platform serving multiple enterprise tenants, models, and security policies. As those services become more complex, isolated infrastructure processing becomes more valuable. Third-party software support can improve the utility of BlueField and strengthen its position in future infrastructure designs. Management’s distinction between GPU rental and token services is particularly important. F5 indicated that bare-metal GPU-as-a-service does not create a similarly significant F5 opportunity, whereas enterprise token services require more governance, isolation, and traffic management. The read-through is therefore stronger for the composition and functionality of AI infrastructure than for aggregate GPU demand. Several boundaries prevent overstatement. Telenor’s selection does not prove a new DPU purchase rather than software deployment on already acquired infrastructure. The disclosed platform uses BlueField-3 and should not be converted into an assumed order for a newer generation. F5’s separate unnamed customer protecting up to 500 models also cannot be identified as NVIDIA from the transcript. The accompanying Telenor announcement describes advanced token-aware governance and inference optimization as planned capabilities. Current selection validates the architecture, but not all prospective utilization or token-cost improvements. These projects were announced before the earnings call, so the call adds context rather than a new award. Near-term trading significance is limited. Subsequent NVIDIA disclosures and sovereign deployment milestones can test whether this architecture spreads, but the call does not support a standalone earnings-estimate increase. The longer-duration implication through 2027 is a larger role for networking, isolation, and security processing as AI factories evolve into commercial services. CYBERSECURITY AND APPLICATION SECURITY PRODUCTION AI VALIDATES AN INDEPENDENT RUNTIME-SECURITY BUDGET (READ-THROUGH 5) Affected companies: Palo Alto Networks (PANW: US) and Cisco Systems (CSCO: US). Direction: positive for AI-security demand, with increasing competitive intensity limiting near-term margin implications. Magnitude: moderate for Palo Alto’s emerging AI-security franchise; smaller at Cisco’s consolidated level. F5 described a production deployment in which an unnamed global AI infrastructure provider selected its guardrails to protect up to 500 models. Requirements included prompt-injection protection, sensitive-data controls, customization, and self-hosted deployment. Management also reported more than 200 active direct AI customers and emphasized enterprises deploying multiple model types across private and public environments. The critical mechanism is that enterprise control requirements extend beyond the safety features of an individual model. Organizations need policy across models, applications, data sources, and agent actions. Those controls can become a separate purchasing requirement when AI moves into production, supporting spending on runtime inspection, model assessment, discovery, and governance. Palo Alto’s Prisma AIRS and Cisco’s AI Defense have verified product exposure to these requirements. Their installed security relationships provide distribution into the relevant buyers, while their runtime products address threats identified on the F5 call. The implication is validation of a real budget category, not proof that either vendor will win F5’s customers. For Palo Alto, this is corroboration of an existing commercial trajectory. Its September earnings call had already disclosed more than $100 million of Prisma AIRS ARR and more than 800 customers. F5’s results provide additional evidence that production AI protection is being purchased across vendors. The figures are not directly comparable: F5 disclosed bookings across multiple AI use cases, predominantly data delivery, whereas Palo Alto disclosed product-specific ARR. The competitive offset is equally clear. F5 plans to double specialist seller headcount, more than double related marketing investment, and incentivize its broader field organization to sell AI security. This increases competition for enterprise budgets and technical resources. A growing market can support multiple vendors without producing rapid margin expansion for all of them. The call also does not establish a security-industry growth explosion. F5’s total security revenue grew 10%, and product security revenue grew 15%, below its overall product growth of 19%. The high-conviction positive concerns the emergence and persistence of production AI-security spending, not an extrapolation of AI customer growth into consolidated cybersecurity revenue growth. Near-term catalysts are the next FY2027 disclosures on paid deployments, AI-security ARR, expansion, and sales productivity. The longer-duration shift is toward model-independent enterprise policy enforcement. That supports security-platform relevance, while leaving vendor share and profitability dependent on execution. AI AGENTS INCREASE THE VALUE OF AUTHENTICATION AND SELECTIVE CONTROL, NOT SIMPLY TRAFFIC VOLUME (READ-THROUGH 6) Affected companies: Akamai Technologies (AKAM: US) and Cloudflare (NET: US). Direction: positive for bot-management, account-protection, and API-security relevance; mixed for immediate monetization. Magnitude: moderate at the product level, with a smaller and unquantified near-term consolidated revenue effect. F5 disclosed that a major financial institution was using Bot Defense to block personal AI agents entering customer accounts, with other enterprises adopting similar protection. Traffic from self-identifying AI agents more than doubled during the quarter. The call also reported an 84% increase in WAF requests and a 103% increase in API-security events. The most actionable implication is not that every additional AI request becomes billable security revenue. It is that identifying whether traffic comes from a human or a bot is becoming insufficient. Enterprises must determine whether an agent is authorized to perform a specific action, on a particular endpoint, with particular credentials. Account access and money movement require different policies from browsing public information. Akamai and Cloudflare have relevant deployed capabilities in bot detection, application security, and traffic-policy enforcement. Akamai’s October 26 discussion of silent agents specifically addressed authenticated account activity and endpoint-level controls. Cloudflare’s recent verified-bot framework likewise distinguishes identity from ongoing behavior. F5’s bank example provides customer-side evidence supporting those already identified product requirements. The initial economic benefit may be retention and product relevance rather than new revenue. F5 explicitly noted that some customers were addressing the problem with capabilities already available to them. Similar functionality at competitors may also be bundled into existing subscriptions. Incremental revenue requires paid upgrades, broader deployment, usage-sensitive pricing, or a distinct new service. There is also a competitive constraint. F5 described a financial-services consolidation win replacing 4 incumbent vendors across application delivery and cloud-related tools. The displaced vendors were not identified, so this cannot be presented as a confirmed Akamai or Cloudflare loss. Nevertheless, it demonstrates that growing security needs can coincide with fewer suppliers and greater pressure on standalone product pricing. The November 5 earnings releases for Akamai and Cloudflare are near-term catalysts for testing security demand, customer expansion, and monetization. The longer-duration positive is that agentic activity makes granular access control more important. The corresponding negative is that neither rising request counts nor greater agent activity automatically produces completed commercial transactions, incremental subscriptions, or improved margins. TELECOMMUNICATIONS AND SOVEREIGN AI SOVEREIGN AI CREATES A COMMERCIAL SERVICE OPPORTUNITY, BUT CAPITAL DEPLOYMENT PRECEDES PROVEN RETURNS (READ-THROUGH 7) Affected companies: Telenor ASA (TEL: Norway) and PT Indosat Tbk, operating as Indosat Ooredoo Hutchison (ISAT: Indonesia). Direction: positive for enterprise-service positioning; negative for early cash requirements where capacity investment precedes utilization. Magnitude: modest near-term group earnings significance, with longer-duration strategic relevance. These are named customers rather than inferred beneficiaries. F5 described Indosat’s Intelligence Suite providing governed access to more than 300 open and closed models, supported by tenant isolation, container security, and GPU-aware load balancing. Telenor selected F5’s service plane for sovereign AI infrastructure serving Norwegian and European enterprises. The transmission mechanism is commercialization of infrastructure that would otherwise be sold primarily as compute capacity. Governance, model routing, tenancy, auditability, and cost controls make shared capacity more usable for regulated enterprise customers. Telecom operators can combine local infrastructure and enterprise relationships with these capabilities to offer a managed service rather than undifferentiated GPU rental. The geographic evidence is supportive. F5’s 31% EMEA growth, 14% APAC growth, and customer examples involving national financial infrastructure indicate demand for local control and consistent policy. This supports the relevance of domestic or regionally governed AI services. It does not establish that every workload must remain local: Indosat’s architecture can route requests to global models where policy permits. The negative cash implication is already visible at Telenor. Its Q2 report identified NOK 150 million of GPU investment in AI Factory and linked that investment to higher group capital intensity. F5’s software selection can improve the platform’s commercial readiness, but it does not demonstrate sufficient utilization, pricing, or customer commitments to earn an attractive return on that capital. The same distinction applies to Indosat. Access to more than 300 models measures platform breadth, not paid demand. A governed interface may increase customer adoption, but model availability cannot be converted into recurring revenue or profit without evidence of customer usage and commercial economics. Contract values and incremental earnings contributions were not disclosed. Telenor’s confirmed October 28 results and Indosat’s upcoming 9-month update provide immediate opportunities to assess commercial progress, investment requirements, and enterprise demand. The F5 partnerships were announced earlier in October, so the earnings call reinforces their significance rather than creating new awards. The longer-duration opportunity is a differentiated enterprise AI service built around local control, governance, and reliable operation. The principal risk is that suppliers monetize equipment and software before operators achieve sufficient utilization. Sovereign AI is a credible new service category, but the call does not yet establish a material improvement in telecom earnings or returns on invested capital.