Daily 2026 · 06 · 04
Economic Tech

DeepSeek opens its first outside round at $59B, Tencent and CATL anchor the China AI capital stack

Hangzhou's DeepSeek is raising about 50 billion yuan, roughly $7.4 billion, in its debut external financing at a valuation between 350 and 400 billion yuan. Founder Liang Wenfeng is committing 20 billion yuan of his own capital. Tencent puts in 10 billion, battery maker CATL 5 billion, NetEase and JD.com 3 billion each, alongside IDG Capital, Monolith, Loyal Valley and Shixiang Tech. The round prices DeepSeek at roughly six times its April mark of about $10 billion and pulls a lab that built its early reputation on cash discipline into the same valuation tier as Anthropic and xAI.

South China Morning Post / CNBC / Manila Times · 2026-06-03 · Source ↗ ·

DeepSeek, the Hangzhou laboratory whose R1 reasoning model rewired the open-weights leaderboard at the end of 2024, is closing the first external financing in its history. South China Morning Post, CNBC and Reuters confirmed the round on June 3 from people familiar with the discussions. The headline number is 50 billion yuan, about $7.4 billion in current FX. The post-money valuation lands between 350 and 400 billion yuan, or $52 billion to $59 billion in dollar terms, depending on the final allocation. Until this raise DeepSeek had funded itself out of the trading profits of High-Flyer, founder Liang Wenfeng's quantitative hedge fund, and had publicly resisted the kind of strategic capital that has become standard at OpenAI, Anthropic and xAI.

Liang is anchoring the round himself. He is committing 20 billion yuan, roughly $3 billion, of his own capital, which keeps him as the single largest shareholder and avoids the dilution path that has made governance friction a recurring story at OpenAI. Tencent Holdings is putting in 10 billion yuan. Battery maker Contemporary Amperex Technology Co. Limited, the CATL of the EV supply chain, is committing 5 billion yuan. NetEase and JD.com each take a 3 billion yuan ticket. The VC layer rounds out the stack: IDG Capital, Monolith, Loyal Valley Capital and Shixiang Tech absorb the balance. None of the named participants is a foreign investor.

From hedge-fund sidecar to national AI champion

DeepSeek's prior valuation as recently as April 2026 was approximately $10 billion. The new mark represents close to a six-fold leap over six weeks. The catalyst is two-sided. On the demand side, DeepSeek's V4 Pro and V4 Flash models, released April 24, hold the top spot on Artificial Analysis's open-weights ranking and lead SWE-Bench Verified among open weights at 77.2. On the supply side, the geopolitical context has changed. The Bureau of Industry and Security's June 1 ruling that closed the China-subsidiary chip-shipment loophole tightened the constraint on US-origin compute reaching Chinese labs, and CATL's participation reads as the energy-infrastructure half of the answer to that constraint.

The CATL position is the most strategically loaded line on the cap table. CATL is the world's largest electric-vehicle battery maker and has spent the past eighteen months pushing into datacenter power, energy storage and reliability equipment. The company's reading of the AI buildout is that the binding constraint shifts from accelerators to power. By taking a direct stake in a frontier-model developer, CATL is buying a strategic option on the demand side of the power-infrastructure trade.

"The line-up underscores China's efforts to build an increasingly self-sufficient AI industry, from models to the energy infrastructure needed to power them."— South China Morning Post, on the investor composition

How the valuation compares

The valuation lands DeepSeek between Anthropic and OpenAI in the open-weights tier. Anthropic's S-1 filed June 1 carried a $965 billion implied value at the IPO price range; OpenAI's pre-IPO secondary marks in May settled around $500 billion. The $59 billion DeepSeek number is closer to the xAI $50 billion April mark and the MiniMax $40 billion June 1 secondary than to the US frontier-lab tier. The comparison the market is making is to the open-weights leaders: DeepSeek's $59 billion sits roughly ten times Alibaba's Qwen team's implied value inside Alibaba Cloud, and roughly fifteen times Moonshot's Kimi-K2.6 implied value.

The round also resolves a long-standing ambiguity about DeepSeek's institutional path. The lab had been described in Chinese trade press as "the High-Flyer team" for most of 2025. The participation of Tencent and JD.com formalizes its independent status as a portfolio company of two of China's largest internet platforms, which gives DeepSeek distribution and capital access without the regulatory headaches of a direct platform subsidiary. The fundraising structure leaves the lab free to ship open weights, which has been the foundation of its growth, while drawing on its strategic investors for cloud distribution and compute access.

What the capital pays for

SCMP and CNBC reporting note the capital will fund three workstreams. The first is the next generation of DeepSeek's training run, sized to compete with GPT-5.5 and Claude Opus 4.8 at the frontier tier rather than only on the open-weights leaderboard. The second is inference infrastructure, including a domestic deployment of DeepSeek's models on Huawei Ascend silicon and possibly on Cambricon and Biren accelerators. The third is the talent and recruiting workstream, with DeepSeek hiring aggressively across robotics, multimodal video and agentic-reasoning research.

What the round does not buy is access to Nvidia frontier silicon at scale. The BIS June 1 clarification closed the China-subsidiary loophole that had been the primary residual channel for restricted-export accelerators to reach Chinese labs through Hong Kong, Singapore and Malaysian-domiciled subsidiaries. DeepSeek will train its next generation primarily on domestic accelerators, which is the technical question the round is most exposed to. The compute-cost discipline that defined V3 and R1 has so far translated to the domestic stack; the question for V5 is whether that translation holds at frontier scale.

"The fundraising could value the company after the investment at between 350 billion yuan and 400 billion yuan."— Reuters, citing people familiar with the matter

Market read-across

Three downstream consequences are already visible. First, China's frontier-lab tier now has a credible third entrant alongside Alibaba's Qwen team and ByteDance's Doubao, both of which are funded internally. Second, the open-weights category has a new floor: a $59 billion lab whose model weights are publicly released. Third, the energy-supply story attached to AI infrastructure now has explicit equity participation from the world's largest battery maker, which is the single clearest signal yet that China's AI buildout will be coupled to its electrification capacity in a way the US story has not yet matched.

The round is expected to close formally before the end of June. Final allocations and the exact post-money number depend on a few late-stage strategic commitments still being negotiated. Liang has not personally commented on the round. Tencent and CATL declined to comment to SCMP, CNBC and Reuters. DeepSeek's press office referred queries to "future communications," the standard formulation the lab has used for press requests since R1.

Big picture

DeepSeek's $59 billion debut round closes the post-R1 valuation gap with the US frontier tier and pulls CATL onto an AI cap table for the first time. Microsoft Build's second-day session put Majorana 2 on stage with a 1,000-fold qubit-reliability gain and a pulled-in 2029 commercial timeline, and shipped Microsoft Discovery to general availability as the agentic-R&D platform behind the same chip. Snowflake Summit ran in parallel with three releases that matter: CoCo as a renamed Cortex Code coding agent with Claude Code and VS Code surfaces, Apache Iceberg v3 in general availability through Horizon Catalog, and Cortex Training for Qwen and Mistral fine-tunes inside the Snowflake boundary.

Anthropic put two consequential pieces of work into the public record in a single morning. The Threat Intelligence team published twelve months of telemetry from 832 banned malicious-cyber accounts, mapped technique-by-technique against MITRE ATT&CK, showing that the share of medium-and-higher-risk actors using Claude rose from 33 percent to 56 percent over the period and that 67 percent used Claude for malware development. The Partner organization restructured the Claude Partner Network with a Services Track ladder of Select, Preferred and Global Premier tiers, plus a Partner Hub portal, against a base of more than 40,000 applicants and 10,000 certified consultants since March.

Venture capital ran wide. Coralogix raised $200 million at a $1.6 billion post-money to build the observability layer underneath autonomous agents, surfacing Olly as a customer-side AI agent embedded in command-line and code interfaces. Apoha exited stealth with $36 million for a wave-form materials AI that claims 90 percent precision on flagging high-risk pharmaceutical candidates pre-clinical. French silicon-quantum startup Quobly closed €115 million from Bpifrance, SEALSQ and STMicroelectronics. Tesla expanded unsupervised Robotaxi to the entire Austin metro on a fleet of 20 active vehicles, against Waymo's 3,000-car operational base, while OpenAI shipped Codex's Computer Use to Windows and Goal mode to general availability across app, IDE and CLI.

More today

18 stories · 2026-06-03
Tech

Microsoft puts Majorana 2 on stage at Build, swaps aluminum for lead, pulls quantum timeline to 2029

Microsoft Source / Neowin · 2026-06-03 · Source ↗ ·

Satya Nadella opened Build day two with Majorana 2, the successor to the Majorana 1 topological processor announced earlier in 2026. The chip's qubit lifetimes jump from milliseconds to an average of 20 seconds, occasionally past a minute, an over-1,000-fold reliability gain. The headline change is materials: aluminum was replaced with lead in the topological stack, more than doubling the gap that protects the qubit from errors. Microsoft now targets 2029 for a scalable quantum computer, half the previous timeline.

Satya Nadella's Build keynote on the morning of June 3 closed with a quantum demonstration that Microsoft has been building toward for the better part of a decade. Majorana 2 is the successor to Majorana 1, which Microsoft introduced in February 2026 as the first topological quantum processor with a working chip rather than a single-qubit test article. Majorana 2 is the first device in the line to combine the topological-qubit architecture with material substitutions that materially extend the time the qubit can hold information.

The headline reliability number is the easiest to remember. Majorana 1's qubit lifetimes were measured in milliseconds. Majorana 2's are measured in seconds, with an average of 20 seconds and occasional measurements past a minute. That is a 1,000-fold reliability improvement on a single generation, larger than any single-generation reliability jump in IBM, Google, Quantinuum, IonQ or Atom Computing's published roadmaps. The number that matters underneath that figure is the topological gap, which is the energy barrier that protects the qubit from the environmental noise that decoheres it. Microsoft says the gap in Majorana 2 is more than double the gap in Majorana 1.

Why lead instead of aluminum

The material change is the load-bearing technical decision. Majorana 1 used aluminum as the superconducting layer in the topological stack. Aluminum was the conventional choice because the field has decades of process knowledge around the metal. Microsoft swapped to lead in Majorana 2. The lead substitution raises the superconducting transition temperature and increases the strength of the topological gap, which translates directly into longer coherence times and lower error rates per logical operation. Lead is more difficult to handle in a semiconductor fab than aluminum, and Microsoft's engineering teams spent the last fifteen months developing the lead-deposition process that makes Majorana 2 manufacturable.

The interface between the lead superconductor and the indium-arsenide-based topological semiconductor is what carries the Majorana modes. The interface chemistry matters because the entire topological-qubit premise rests on a clean, atomically reproducible interface between the two layers. Microsoft has said that yield on the lead-process Majorana 2 wafers is the bottleneck that determines how quickly the line scales from current single-chip devices to the multi-chip arrays that a useful quantum computer requires.

"Majorana 2 represents a huge step towards the goal of a million qubits on a single chip that can fit in the palm of the hand."— Satya Nadella, Microsoft Build keynote, June 3, 2026

The pulled-in commercial timeline

The headline corporate consequence is that Microsoft has cut its 2030s scalable-quantum-computer target in half. The new target is 2029. The 2029 date is the first time Microsoft has named a delivery year for a scalable system rather than a research milestone. Microsoft's quantum roadmap, presented at Build by Krysta Svore and Chetan Nayak, breaks the path into three milestones: a 100-qubit array by mid-2027, a 1,000-qubit array by late 2028, and a 1 million qubit array by 2029. The million-qubit milestone is the threshold at which quantum advantage on industrially relevant problems becomes achievable; Microsoft's argument is that the topological architecture's intrinsic error suppression collapses the error-correction overhead that has dragged the IBM and Google roadmaps past 2030.

The Build presentation was paired with a Microsoft Discovery announcement (covered separately in today's edition) that disclosed which agentic-AI workflows had been used to design Majorana 2. The chip is the first major Microsoft research output to attribute material design decisions to an agentic-AI workflow rather than to human researcher intuition alone. Microsoft Discovery's role was in the screening of candidate superconductor-semiconductor material pairs, of which the lead-on-InAs combination emerged as the strongest design candidate.

Where the field stands

The competitive frame against the rest of the quantum field is the gap-versus-overhead trade. IBM, Google, Quantinuum and Atom Computing are pursuing superconducting and trapped-ion architectures that require thousands of physical qubits to construct a single logical qubit because the underlying error rates are too high. Microsoft's topological qubits, if the architecture holds, require closer to a 1-to-1 physical-to-logical ratio. The Majorana 2 reliability number is the strongest evidence to date that the topological premise is sound at chip scale, not just at single-qubit test scale. Independent verification by external groups will follow the Microsoft Research paper Microsoft is expected to submit to Nature in Q3.

TechApplication

Microsoft Discovery hits general availability with a desktop app preview, an evidence-to-hypothesis loop, and Majorana 2 as flagship reference

Microsoft Azure Blog / AIwire · 2026-06-03 · Source ↗ ·

Microsoft Discovery shipped to general availability at Build day two as an agentic R&D platform with the Microsoft Discovery Engine at its center. The platform lets organizations build specialized agents, connect them to institutional knowledge and external scientific corpora, and run them across modeling, simulation, analysis and validation. Microsoft also opened a local desktop app preview for researchers. The Majorana 2 material-screening workflow is the first published flagship reference customer.

Microsoft Discovery moved from a closed preview, where it had run since the Build 2025 introduction, to general availability for all organizations on the morning of June 3. The launch material frames Discovery as an "agentic R&D platform" rather than as another LLM endpoint. The pitch is that scientific work is structured as a loop: evidence to hypothesis, hypothesis to experimental design, design to execution, execution to analysis, and analysis back into the next iteration. Discovery is the orchestration substrate that holds the loop together and lets research teams operate on it through a shared system of agents and knowledge.

The platform centers on a component Microsoft calls the Microsoft Discovery Engine. The Engine handles the loop's state, manages the connections between agents and the external tools they need to run modeling, simulation, analysis and validation work, and surfaces the resulting outputs back into the institutional knowledge base. Discovery is designed as an enterprise platform; the governance and access-control layer is treated as load-bearing rather than as an afterthought. The platform exposes the standard Microsoft enterprise controls around identity, data residency, audit and key management through Azure Foundry's existing primitives.

What ships at GA

Discovery's agent model lets teams instantiate specialized agents for individual research workflows, then have those agents call into modeling and simulation tools, retrieval over institutional and external knowledge, and human-in-the-loop validation steps. The platform handles parallel execution across HPC backends and the orchestration of multi-step experimental designs. Microsoft also introduced the Microsoft Discovery app in preview, which is a local desktop experience targeted at researchers, students and smaller scientific teams that want to work with the platform without managing the full Azure Foundry deployment. The desktop app is available on Windows and macOS in private preview through the Microsoft Foundry portal.

Pricing is consumption-based through Azure. Microsoft did not disclose a single per-token or per-agent rate at GA but tied billing to the underlying Foundry-resource consumption: compute hours for HPC simulation, token spend for the agent layer, and Foundry endpoint costs for the model calls.

"Microsoft Discovery brings together agentic orchestration, advanced reasoning, knowledge foundations, and high-performance computing to modernize scientific discovery across biology, materials and engineering."— Microsoft Azure announcement post, June 3, 2026

Majorana 2 as the reference workflow

The launch is paired with a published case study on the role Discovery played in designing the Majorana 2 quantum chip Nadella unveiled in the same keynote. Discovery agents ran a virtual screening pipeline across superconductor-semiconductor pairings, scored candidate materials against criteria for topological-gap strength, fab-process tractability and interface stability, and produced a ranked candidate list. The lead-on-indium-arsenide pairing that became Majorana 2's material stack emerged from this screening. The story is the first instance of Microsoft documenting an agentic-AI workflow as the proximate cause of a hardware-engineering decision rather than as a productivity tool sitting alongside human researchers.

The competitive frame

Discovery enters a category Google has been building under its AlphaFold-3 / Lab Manager program, Anthropic has been seeding through the Mythos Glasswing safety-oriented research line, and OpenAI has been pursuing through its biology and materials-science partnerships announced in April. Microsoft's positioning leans on the Azure Foundry integration: customers already running enterprise AI on Foundry can extend their existing governance, identity and data-residency posture into the Discovery agent stack without re-doing security review. The Mayo Clinic frontier-health partnership announced June 2 was framed in the keynote as an early Discovery customer; June 3's launch makes the platform broadly available to the rest of the customer base.

SafetyTech

Anthropic publishes a year of AI-cyber telemetry, share of medium-risk actors rises 33% to 56%, 67% used Claude to build malware

Anthropic Threat Intelligence · 2026-06-03 · Source ↗ ·

Anthropic's Threat Intelligence team published a twelve-month analysis covering March 2025 through March 2026, mapping 832 banned malicious-cyber accounts technique-by-technique against MITRE ATT&CK. Across those accounts, researchers logged 13,873 observed actions and 482 unique techniques. Medium-or-higher-risk actors rose from 33 percent in the first half to 56 percent in the second half. Anthropic shared the findings with Verizon for inclusion in the 2026 Data Breach Investigations Report.

Anthropic's Threat Intelligence team published the first systematic, year-long telemetry analysis of how malicious actors are using a frontier model. The report covers March 2025 through March 2026 and draws on the 832 accounts the company banned for malicious-cyber violations during the period. The team mapped each ban against the MITRE ATT&CK adversary-technique framework, the de facto reference taxonomy that security analysts use to describe what an attacker did, how they did it, and which defensive controls would have caught them. The output is 13,873 observed actions and 482 unique techniques inside Claude conversations.

The headline statistic is the share of banned actors classified as medium or higher cyber risk. In the first six months of the window, the share was 33 percent. In the second six months, the share was 56 percent. That is a 1.7-fold increase over twelve months. The interpretation Anthropic puts on the number is that the actor population using Claude for malicious cyber work has shifted from low-skilled experimenters toward more sophisticated operators in a single year. The team also notes that the per-actor average technique count rose: the least-skilled cohort averaged 16 techniques and the most-skilled cohort averaged 20, but the gap between novice and sophisticated actors shrank because LLMs are flattening the skill ceiling.

Where the LLM lift is concentrated

Malware development is the single most common use case. About 67 percent of analyzed accounts, roughly 560 accounts in absolute terms, used Claude to help write malicious software. The second-largest category is account discovery and reconnaissance work, which rose by 8.9 percentage points across the window. AI-assisted phishing declined by 8.6 points, which the report attributes to phishing's lower technical barrier; LLM assistance offers less marginal lift on a task that does not require novel code. The smallest individually-detected category is lateral movement, which still represented 6.5 percent (54 accounts) of the dataset.

The November 2025 state-sponsored cyber-espionage operation Anthropic disrupted is the central case study. The operation used 30 MITRE techniques across 13 tactics and required minimal human intervention; the AI orchestrated the post-compromise chain autonomously. Anthropic's argument is that this case marks an inflection in what the threat-intelligence community calls the post-compromise lift problem: the operations that traditionally separated nation-state actors from less-resourced groups, lateral movement, privilege escalation, persistence and exfiltration, are exactly the steps a capable LLM-orchestrated agent can now run end-to-end without recourse to human expert tradecraft.

"Cybersecurity has reached a critical inflection point, with the percentage of medium-or-higher-risk actors rising from 33 percent to 56 percent over a single year."— Anthropic Threat Intelligence report, June 3, 2026

Methodological notes and limits

The 832 accounts are a subset of the total banned for malicious cyber activity in the window. Anthropic restricted the dataset to cases where the available telemetry was rich enough to support technique-level mapping. The MITRE ATT&CK framework is the reference taxonomy, but the report identifies a gap: the existing framework's adversary-technique vocabulary does not have first-class terms for agentic orchestration or for the autonomous chaining of attack stages that the November espionage case illustrates. Anthropic is engaged with MITRE on extending the framework to capture these patterns and is publishing the underlying technique mapping as a contribution to that workstream.

The report also flags a measurement caveat. The 832 banned accounts are accounts Anthropic detected. The 1.7-fold rise in medium-risk share may reflect either a real population shift or improved detection, or both. The team's view is that the detection improvements are real but not large enough to account for the entire shift; the population is changing.

Industry sharing

Anthropic shared the findings with Verizon for inclusion in the 2026 Data Breach Investigations Report, the de facto industry reference, published in late May. Verizon's recognition of the Anthropic Safeguards Team named Kyla Guru and Jacob Klein as research partners. The DBIR's headline finding for the period, that vulnerability exploitation overtook credential theft as the top breach vector, is consistent with the Anthropic shift toward more sophisticated actors using AI to scale the post-compromise lift.

"AI enables attackers to perform complex post-compromise operations that traditionally required expert knowledge."— Anthropic, on the post-compromise lift problem

The defensive recommendation Anthropic underscores is that traditional risk-assessment signals, like technique count and platform type, no longer correlate reliably with threat level. The recommendation is to lean harder on behavioral and chain-level detection, which is the workstream Anthropic's classifier system is built around. The full technique mapping is published in machine-readable form for use by security vendors and CIRTs.

EconomicApplication

Anthropic restructures the Claude Partner Network with a Services Track ladder and a Partner Hub portal

Anthropic · 2026-06-03 · Source ↗ ·

Anthropic formalized its enterprise channel on June 3 with a three-tier Services Track. Select needs 10 certified individuals, 2 production deployments and 1 public reference. Preferred raises the bar to 100 certified, 15 deployments and 3 references. Global Premier sits at 1,000 certified, 100 deployments across three regions, 15 references and a joint business plan. Promotions process on January 1 and July 1 with an additional October 1 review in 2026. Forty thousand firms have applied since March and 10,000 consultants are certified.

The Claude Partner Network was launched in March with a $100 million Anthropic investment behind training, technical support and shared marketing. The structure was deliberately light: an application, a certification path, and a small set of program benefits. Three months in, the program's volume has outrun the structure: 40,000 firms have applied and 10,000 consultants have completed certification. The June 3 announcement formalizes the network into a tiered Services Track with three rungs and adds a Partner Hub portal that both partners and customers can use to see where each firm stands against the requirements.

The three tiers and the bar at each

Select is the entry tier. The requirements are 10 or more active certified individuals, at least 2 joint customers running Claude in production over a trailing 12-month window, and at least 1 public customer story. The threshold is set low enough that most firms with serious early Claude practices qualify, and the program design treats Select as the on-ramp from which Preferred and Premier flow.

Preferred is the middle tier. The bar lifts to 100 certified individuals, 15 deployed customers, and 3 published stories. The tier targets the firms that have built named Claude practices with five-to-ten consultants per client and multi-product delivery experience. Global Premier is the top of the ladder. The requirements are 1,000 active certified individuals, 100 deployed customers spanning at least three regions, 15 public customer stories, and a joint business plan with named executive sponsors at both Anthropic and the partner firm. The 1,000-consultant requirement is calibrated to the staffing scale of the largest global-systems-integrator practices.

"Promotions are processed twice a year, on January 1 and July 1, with an additional review on October 1, 2026, in this first year."— Anthropic Services Track announcement

Who is already in

The launch material names the largest global-systems-integrator partners and the size of their Claude-trained workforces. Accenture has 30,000 professionals trained. Cognizant covers approximately 350,000 associates. Deloitte's number is 470,000. KPMG sits at more than 276,000. PwC is described as covering "hundreds of thousands" with global expansion underway. The named partners overlap heavily with the existing Microsoft, Google and AWS partner ecosystems, but Anthropic's pitch is that the Claude practices are growing as a distinct line of business at each firm rather than absorbing existing AWS-resold-Claude revenue.

The Partner Hub portal

Partner Hub is the second piece of the announcement. The portal is a two-sided platform. On the partner side, firms see exactly where they stand against the tier thresholds, including which certifications are open or expiring, which deployments count toward the deployed-customer requirement, and which public stories have been validated. On the customer side, a buyer browsing for a Claude practice can filter on tier, deployment industry, regional coverage and named customer references, and short-list firms qualified for the project scope. The portal is the formal mechanism by which Anthropic turns an opaque partner directory into an underwritten qualification claim.

The cadence and downgrade mechanics

Promotions are processed twice a year, on January 1 and July 1. An additional review on October 1, 2026, was added specifically for the first year of the new program to handle the existing partner base's qualification. Annual downward reviews occur on December 31 with 90-day notice, which gives firms a window to recover the criteria they slipped against before the demotion takes effect. The structure is closer to the AWS Partner Network's tier mechanics than to the looser Microsoft or Google equivalents and signals Anthropic's intent to run the program as an enterprise channel rather than as a marketing layer.

TechApplication

Snowflake Summit closes day two with CoCo agent, Iceberg v3 GA, Datastream Kafka service and Cortex Training for Qwen and Mistral

Snowflake Newsroom / SiliconANGLE · 2026-06-03 · Source ↗ ·

Snowflake renamed Cortex Code to CoCo and broadened the agent's surface to desktop and mobile apps, Salesforce Slack, Excel, VS Code and Anthropic's Claude Code. Apache Iceberg v3 reached general availability through Horizon Catalog with bi-directional read and write against external engines through Apache Polaris. Snowflake Datastream is in private preview as a managed Kafka service. Cortex Training opens managed-GPU fine-tuning of open-source Qwen and Mistral inside the Snowflake boundary.

Snowflake Summit's second day on June 3 closed with four releases that consolidate the company's pivot from data-warehouse vendor to enterprise-AI control plane. The headline release is CoCo, the rebrand of Cortex Code as Snowflake's enterprise coding agent. CoCo's surface expanded from the original Snowflake-only UX into a multi-environment delivery model: desktop and mobile apps, integration with Salesforce-owned Slack, integration with Microsoft Excel, integration with Visual Studio Code, and explicit interoperability with Anthropic's Claude Code. The Claude Code piece is the load-bearing line on the partnership side; it makes Snowflake's data plane a first-class destination for the agentic coding workflow Claude Code drives.

Apache Iceberg v3

The Apache Iceberg v3 general-availability announcement is the largest of the data-platform releases. Iceberg v3 has been Snowflake's beta channel since the November 2025 introduction; the GA flips the format on by default for Snowflake-managed tables. Horizon Catalog, the metadata layer Snowflake repositioned around Apache Polaris earlier in 2026, now provides bi-directional read-and-write access to Snowflake-managed Iceberg tables from external compute engines using open standards. The practical implication is that a Snowflake customer's Iceberg tables can be queried and written by Databricks, BigQuery, Trino, Presto, DuckDB or any other Polaris-compatible engine without the customer needing to maintain dual catalogs.

The interoperability is the point. Snowflake's pitch through the day was that the data layer should be neutral with respect to which agent or engine touches it. The metadata, lineage, identity, policy enforcement and business context travel with the data through Polaris, not with the engine that originally created the data. That positioning is a direct response to the parallel agent-runtime announcements from Databricks at Data + AI Summit in two weeks.

Datastream and Cortex Training

Snowflake Datastream is a managed streaming service for Apache Kafka, in private preview at GA-time. The service combines real-time data ingestion with the AI pipeline in a single Snowflake-billed plane. The use case is the streaming-to-agent workflow that has been the hardest piece of enterprise AI to operationalize: Datastream lets a customer wire a Kafka topic into a Cortex agent without standing up a separate Kafka cluster and without exporting data outside the Snowflake compliance boundary.

Cortex Training is the fourth release. The service provides managed GPU infrastructure inside Snowflake's perimeter that lets customers fine-tune open-source Qwen and Mistral models against their proprietary data without removing the data from Snowflake. The service is targeted at the enterprise governance problem: most regulated industries cannot move training data to a third-party fine-tuning vendor, and Cortex Training collapses the workflow into a single billed service whose data plane is the Snowflake data plane. The launch list of supported base models is Qwen and Mistral at GA, with explicit hooks to extend the catalog as new open-weights models clear Snowflake's security review.

"Metadata, lineage, identity, policy enforcement and business context should travel with the agent, not stay locked up in the platform it started in."— Snowflake Summit positioning, agentic-enterprise track

The strategic frame

Snowflake is competing for the same enterprise-AI control-plane opportunity as Databricks, Microsoft Fabric, Google Cloud's Vertex AI Studio and SAP's Joule. The differentiation Snowflake leans on is the data-governance posture: by treating Iceberg, Polaris and the Cortex stack as one continuous control plane, the company is positioning the data layer as the durable enterprise asset against which agents come and go. CoCo's interoperability with Claude Code, rather than competition against it, is the clearest illustration of the bet. Snowflake is not trying to win the agent layer outright. It is trying to be the data destination every winning agent connects to.

Pricing was disclosed in line with the rest of Snowflake's consumption model. CoCo billing follows the Snowflake credit consumption profile; Cortex Training is priced against GPU hours with a separate base-model licensing add-on; Datastream is billed against stream throughput and storage. The Iceberg v3 transition does not change billing for Snowflake-managed Iceberg tables, but external-engine access through Polaris incurs the standard Horizon Catalog metadata fee.

ApplicationTech

Tesla opens unsupervised Robotaxi to the entire Austin metro on 20 active cars, fleet has shrunk from 165 to 34

Electrek · 2026-06-03 · Source ↗ ·

Tesla's official Robotaxi account posted on the morning of June 3 that the service now covers the entire Austin metro area. The unsupervised fleet stands at about 20 active vehicles, down from a late-April peak of 25, and the total Tesla autonomous fleet across all markets including supervised Bay Area has collapsed from 165 to 34. Dallas sits at 3 vehicles, Houston at 6, neither has added a car since April launch. Waymo is at roughly 3,000 vehicles and 500,000 paid trips a week.

Tesla's official Robotaxi X account posted the announcement on the morning of June 3: "Unsupervised Robotaxi now in the entire Austin Metro area." The geographic expansion is real and large. The Austin metro encompasses several thousand square miles, including suburbs like Round Rock, Cedar Park, Pflugerville and Georgetown. The previous boundary had been a small geofence in South Austin around the original launch zone. The new boundary is roughly the same scope of operational design domain that Waymo announced for Austin earlier in 2026, before Waymo paused its Austin expansion.

The fleet behind the expansion is the point of friction. Tesla currently fields about 20 active unsupervised Robotaxi vehicles in Austin. The unsupervised count peaked at 25 in late April and has drifted down since. Dallas has 3 unsupervised vehicles. Houston has 6. Neither Texas market has added a vehicle since the April launch in those cities. Tesla's overall fleet across all Robotaxi markets, including the supervised Bay Area operation that uses safety operators rather than driverless vehicles, has fallen from 165 to 34 active vehicles. The shrinkage reflects vehicles being pulled offline for software updates and the cadence of FSD v15 development, the rewrite Elon Musk has said must ship before Tesla scales the fleet meaningfully.

The Waymo comparison

Waymo's operational base sits at approximately 3,000 vehicles across 11 US cities. The company reports more than 500,000 paid trips per week. Waymo's geographic coverage as measured by operational-design-domain area is 1,400-plus square miles in active commercial operation across San Francisco, Los Angeles, Phoenix, Austin, Atlanta, Miami, Dallas, Denver, Boston, and others. Waymo raised $16 billion earlier in 2026 specifically to fund the expansion of the operational fleet to support new market launches. The gap between Tesla's 20-car Austin fleet and Waymo's nationwide 3,000-car operation is the structural fact the June 3 expansion does not change.

"Tesla is waiting for its FSD v15 software rewrite before scaling the fleet aggressively."— Electrek, on the rationale for the fleet pause

What the announcement signals

The geographic expansion is plausibly a marketing move ahead of the second-half Robotaxi product rollout Tesla telegraphed at its Q1 earnings call. With FSD v15 still in development, the marginal cost of expanding the operational geography on the existing 20-vehicle fleet is low, and the marginal signaling value to investors and regulators is non-trivial. The risk is real-world wait times: if the 20-car fleet has to cover the entire Austin metro, the per-ride wait time during demand peaks will exceed what a Waymo or Uber rider considers acceptable. Tesla has not published wait-time telemetry from Austin.

The announcement also lands during a period of intense competitive pressure on Tesla. Waymo's purpose-built Ojai robotaxi, unveiled June 2, is the first generation of Waymo hardware designed specifically as a passenger vehicle rather than as a modified production car; it includes the sixth-generation Waymo autonomous-driving stack with cold-weather sensor capability. The Tesla Austin announcement extends Tesla's geographic claim while the competitor extends its hardware claim.

EconomicApplication

Coralogix raises $200M Series F at $1.6B to monitor the AI-agent runtime, Olly agent runs the customer side

TechCrunch · 2026-06-03 · Source ↗ ·

Coralogix, founded in Tel Aviv in 2014 and now headquartered in Boston, closed a $200 million Series F led by Advent and the Canada Pension Plan Investment Board, with Greenfield Partners and Brighton Park Capital. The round prices the company at $1.6 billion post-money and brings total funding to $550 million. Revenue grew more than 60 percent year-over-year and Coralogix counts roughly 30 enterprise customers at $1 million-plus annual spend. The Olly AI agent is at the center of the pitch: more than half of enterprise customers interact with Coralogix through Olly or their own model.

Coralogix raised a $200 million Series F at a $1.6 billion post-money valuation. The round was led by Advent and Canada Pension Plan Investment Board, with participation from Greenfield Partners and Brighton Park Capital. The deal brings Coralogix's total funding to $550 million and is the company's largest round since its 2014 founding. Coralogix is headquartered in Boston, was founded in Tel Aviv by Ariel Assaraf and Yoni Farin, and employs more than 600 people globally with about 100 in India.

The investment thesis is the rise of autonomous AI agents as a class of system that has to be monitored differently than traditional software. The premise is that an AI agent is not a deterministic process whose execution can be traced through structured logs and metrics. It is a probabilistic process whose behavior must be observed and explained in terms a human operator can act on. Coralogix's pitch is that the existing observability vendors, Datadog, New Relic and Splunk, were built around the deterministic process model and that the agent-runtime monitoring opportunity is large enough to support a new category leader.

Olly and the agentic UX

Olly is Coralogix's customer-side AI agent. The pitch is that operators no longer interact with the observability platform through dashboards; they interact with Olly or with their own command-line and agentic interfaces, and Olly handles the translation between human intent and the underlying telemetry system. Coralogix reports that more than half of its enterprise customer base now interacts with the platform through Olly or through customer-owned AI agents that call into Coralogix through APIs. The CEO Assaraf described the underlying shift as "the interface layer is slowly getting eroded."

"The interface layer is slowly getting eroded."— Ariel Assaraf, Coralogix co-founder and CEO

Revenue and customer base

Coralogix's headline financial metrics: revenue grew more than 60 percent year-over-year; the company surpassed $100 million in annualized revenue more than a year ago and is now well above that mark; approximately 30 enterprise customers spend $1 million-plus annually; total customer count is more than 5,000 worldwide. Named customers include IBM, Tradeweb and JFrog. Coralogix says it expects to reach profitability over the next few years; the Series F is sized to extend the runway into the agent-runtime opportunity without forcing a profitability inflection on the current operating model.

The competitive frame

The Series F arrives less than a year after Coralogix's previous raise. The cadence reflects the size of the agent-monitoring opportunity the investors see and the urgency of getting in ahead of Datadog and Splunk extending their offerings. Datadog announced its agentic-observability beta in April; Splunk shipped a corresponding capability in the Cisco Splunk integration in May. Coralogix's distinctive bet is the user-experience layer: Olly as a customer-side agent against the dashboard-driven UX of its competitors.

TechEconomic

Apoha exits stealth with $36M Series A, claims 90 percent precision flagging high-risk drug candidates pre-clinical

Fortune · 2026-06-03 · Source ↗ ·

London- and San Francisco-based Apoha emerged from stealth on June 3 with a $36 million Series A, led by Singular with Draper Associates, Redalpine, Seedcamp, Wilbe and Nucleus, plus an Innovate UK grant. The company trains AI on "wave-form" data from materials suspended in liquid under physical excitation, a measurement system the founders call VIBE. Pre-clinical drug screening is the lead application; lipid nanoparticles, plant proteins, food formulation and durability testing are on the roadmap. Founders are Shamit Shrivastava and Anshika Srivastava.

Apoha emerged from stealth on June 3 with $36 million raised across a Series A and earlier seed tranches. Singular, the European venture-capital firm, led the Series A. Draper Associates participated alongside existing seed backers Redalpine, Seedcamp, Wilbe and Nucleus. An Innovate UK grant rounds out the funding stack. The company is dual-headquartered in London and San Francisco. The founders are CEO Shamit Shrivastava, a mechanical engineer with a PhD from Boston University and postdoctoral work at Oxford, and COO Anshika Srivastava, a former Goldman Sachs banker. The headcount and the founders' technical training are the lines investors cited.

The data and the measurement system

The technical premise is "wave-form" data captured from materials suspended in liquid and excited under controlled physical stress. The company calls the measurement platform VIBE, for Variations in Inter-facial Behavior Under Excitation. The output is a high-dimensional time-series signature unique to each material under each set of excitation parameters; the AI models are trained on the relationship between these signatures and downstream material behavior. The company describes the approach as "liquid intelligence," a positioning device the founders use to distinguish the modality from the existing structure-based and language-based modalities in computational materials science.

"Machines have learned to see what matter looks like. They have not learned to taste, smell, or feel matter."— Anshika Srivastava, Apoha COO

Applications and the demonstration

The lead commercial application is pre-clinical pharmaceutical screening. Apoha says its models, trained against historical pre-clinical failure data, identify high-risk drug candidates with greater than 90 percent precision. The headline application is the dispositioning of candidates before they enter expensive pre-clinical and Phase 1 studies, which is the financial bottleneck pharmaceutical R&D budgets are most exposed to. The same modality extends to lipid-nanoparticle behavior prediction (relevant for mRNA delivery), plant-based protein formulation, food and beverage product development, and materials durability testing.

The model architecture is not disclosed in the launch material. Apoha says the models are proprietary and that the moat is the combination of the VIBE measurement system, the underlying training data, and the model architecture together, rather than any single element on its own. The company is targeting pharmaceutical and food-and-beverage enterprise customers as the first revenue cohorts, with materials-science customers further down the roadmap.

The technical context

Apoha sits in the same broad space as the structure-based AlphaFold-3 / RoseTTAFold lineage and the property-prediction work that has come out of MIT, Stanford and Berkeley over the past three years. The novel claim is the modality: AlphaFold-3 and its successors work from sequence and structure; Apoha works from a time-series physical-excitation signature that the company argues captures information neither sequence nor structure alone can encode. The argument is empirical rather than theoretical: the company is selling the pre-clinical-screening precision number as the proof, and the rest of the model architecture as the trade secret behind the precision.

The $36 million Series A is intended to fund the pre-clinical screening commercial rollout, the expansion of the VIBE measurement platform to additional material classes, and the team build-out across London and San Francisco. The company has not disclosed a customer list at launch.

EconomicTech

Quobly closes €115M Series A to push silicon-based quantum computers out of the Grenoble lab

PR Newswire / Tech.eu / Sifted · 2026-06-03 · Source ↗ ·

Grenoble-based Quobly raised €115 million in Series A funding, led by Bpifrance, SEALSQ and STMicroelectronics, with participation from the European Innovation Council, Blast, ALIAD (Air Liquide Venture Capital) and existing investor Innovacom. The financing routes through Bpifrance's Deep Tech 2030 fund under the France 2030 initiative. Quobly's qubits use 300-millimeter FD-SOI silicon manufacturing. A first commercial cloud-accessible system ships under the Alloy product line by end of 2026, with HPC deployment in 2027.

Quobly, the silicon-quantum-computing company spun out of the CEA-Leti lab in Grenoble in 2022, closed a €115 million Series A on June 3 led by Bpifrance, SEALSQ and STMicroelectronics. Existing investor Innovacom returned alongside the European Innovation Council, Blast and Air Liquide's venture capital arm ALIAD. Bpifrance's participation routes through its Deep Tech 2030 fund, which Bpifrance manages on behalf of the French government as part of the France 2030 industrial-policy initiative. The €115 million is the largest single Series A in European quantum computing to date.

The technical approach is the load-bearing differentiator. Quobly's qubits are silicon-based and fabricated on the FD-SOI 300-millimeter wafer process the European semiconductor industry already runs at industrial scale. The premise is that quantum-computing scale-up will not happen on the bespoke fab processes the superconducting and trapped-ion competitors require, but on the established CMOS-compatible processes that the existing semiconductor supply chain knows how to manufacture at volume. The argument is that silicon qubits trade some single-qubit performance for the manufacturability and yield curve that conventional silicon enjoys, and that the trade is the right one for industrial-scale deployment.

The Alloy product line and the commercial timeline

Quobly plans to ship its first commercial quantum computer through the cloud by the end of 2026 under its Alloy product line. The cloud-accessible system is the first commercial deployment milestone. A 2027 deployment inside HPC infrastructures, sold to scientific-computing customers running coupled classical-quantum workflows, is the second milestone. The company is headquartered in Grenoble with offices in Singapore and Canada and employs more than 100 people. The timeline is aggressive relative to the IBM, Google and Atom Computing public roadmaps, which target 2028 and beyond for the same scale of accessible deployment.

The strategic frame in European industrial policy

The investor list is the European industrial-policy ecosystem in miniature. Bpifrance is the French sovereign investor. STMicroelectronics is the European pure-play semiconductor manufacturer with FD-SOI scale. SEALSQ is the post-quantum cryptography subsidiary of WISeKey that has been the financial anchor of several quantum-related raises in 2026. ALIAD brings industrial-gases expertise relevant to the cryogenic-cooling supply chain. The Series A is sized to fund the Alloy cloud system, the team expansion, and the early HPC-customer deployment work; the structure leaves room for a Series B in late 2027 or 2028 if Alloy reaches the cloud-deployment milestone on schedule.

The Microsoft Majorana 2 announcement on the same day from Build 2026 supplies the competitive context. Microsoft's topological-qubit approach is at the high-risk-high-reward end of the field; Quobly's silicon-qubit approach is at the lower-risk-incremental-yield end. The two architectures will likely coexist commercially for years before the field consolidates on a winner.

ApplicationTech

OpenAI extends Codex Computer Use to Windows apps, ships Goal mode to GA, opens Sites preview

OpenAI Codex Changelog · 2026-06-03 · Source ↗ ·

OpenAI's Codex changelog shipped three updates on June 3. Computer Use now supports Windows apps inside the Codex app, letting eligible users see, click and type in Windows applications while Codex tests and refines builds. Goal mode reached general availability across the Codex app, IDE extension and CLI, replacing the prompt-by-prompt UX with an outcome-and-success-criteria framing. Sites entered preview as a Codex-hosted way to deploy websites, dashboards, internal tools, web apps and games on OpenAI infrastructure.

The OpenAI Codex changelog published three updates on June 3 that together shift the Codex product from a code-generation surface into a goal-driven build-and-deploy environment. The most consequential update is the addition of Windows app support to Computer Use inside the Codex app. Computer Use, which has shipped progressively through Q1 and Q2 of 2026 for macOS and for browser-based desktops, now lets Codex see, click and type inside Windows applications while testing and refining the code it has been asked to build. The Windows extension closes the gap with Anthropic's Claude Computer Use, which has had Windows support since the April 2026 update.

Goal mode and Sites

Goal mode is the second update and the most architecturally meaningful. Goal mode reached general availability across the Codex app, the IDE extension and the command-line interface. The mode changes the user-developer interaction from a turn-by-turn prompting loop into a goal specification: the user defines an outcome and a set of success criteria, and Codex keeps working toward the outcome through iterative refinement, testing and self-correction until the criteria are satisfied or the agent surfaces a blocker. Goal mode in the IDE extension is positioned as a direct response to Cursor's Background Agents and to GitHub Copilot Workspace's autonomous-build mode that reached GA at Microsoft Build the previous day.

Sites is the third update and the most product-distinct. Sites entered preview in the Codex app as a Codex-hosted deployment surface. The use case is the end-to-end workflow from a goal-mode build to a deployed website, dashboard, internal tool, web app or game without leaving the Codex environment. The deployment target is OpenAI-hosted infrastructure rather than the customer's own cloud; the preview is positioned for prototyping and internal tools rather than for production workloads.

The competitive frame

The three updates land in the same week as the GitHub Copilot Workspace GA announced at Microsoft Build day one, the Cursor Teams pricing restructure, and the Windsurf rebrand to Devin Desktop. The agentic-coding category is consolidating into a small number of full-stack offerings that span code generation, test execution, computer-use validation and now deployment. OpenAI's positioning is that Codex is the OpenAI-native end-to-end environment for the same workflow Microsoft is building inside Copilot and Anthropic is building inside Claude Code. The three June 3 updates together close the most visible feature gaps that had separated Codex from its competitors heading into Build.

Pricing for the new capabilities follows the existing Codex profile metered against ChatGPT plan credits. Computer Use on Windows is available to eligible users on the Codex app's higher plan tiers; Goal mode is available across all paid Codex surfaces; Sites is in preview without a separate billing line, with the expectation that deployment-resource consumption will be billed once the preview moves to GA.

Tech

Nvidia and TSMC put AI inside the fab, agents drive design, yield and energy across the manufacturing lifecycle

NVIDIA GTC Taipei / Electronics360 · 2026-06-03 · Source ↗ ·

Day three of NVIDIA GTC Taipei at Computex closed with TSMC committing to use NVIDIA AI and computing chips across the semiconductor design and manufacturing lifecycle. The aim is improvement on turnaround time, energy efficiency, yield and operational productivity inside the fab itself. NVIDIA also previewed AI for Media building blocks for live production, localization, content analysis and synthetic-video detection, alongside DLSS 4.5 Ray Reconstruction and expanded RTX Spark availability.

NVIDIA GTC Taipei wrapped its third day at Computex 2026 with a partnership announcement that puts the fab itself on the list of NVIDIA's AI customers. TSMC is committing to use NVIDIA AI and computing chips across the semiconductor design and manufacturing lifecycle. The four metrics named in the announcement are turnaround time, energy efficiency, yield and operational productivity. The partnership covers TSMC's design-rule compliance and physical verification workflows, the photolithography optimization pipeline, the in-line metrology and defect-detection workstream, and the equipment-uptime and energy-use optimization that determine fab operating cost.

The semiconductor industry has been using AI inside the fab in production for several years, but the TSMC-NVIDIA partnership is the broadest publicly-disclosed commitment to date. The economic logic is straightforward: a 1 percent yield improvement on TSMC's 2-nanometer process represents several billion dollars of annual gross profit; a 1 percent reduction in turnaround time on the fab's most loaded production lines represents a meaningful uplift in TSMC's total addressable revenue at current capacity utilization. NVIDIA's AI accelerators are the highest-yield, highest-margin product TSMC manufactures, and the partnership creates a tight loop in which NVIDIA's customers help make the silicon NVIDIA sells back to them.

The AI for Media announcements

NVIDIA also previewed a new set of AI for Media technologies aimed at broadcasters, streaming platforms and developers. The building-block list covers real-time live production, localization, content analysis and synthetic-video detection. The synthetic-video-detection piece is the most strategically loaded: the announcement positions NVIDIA's media stack as the substrate underneath the verification work newsrooms and streaming platforms increasingly have to do as generative-video tools become widely available.

The Computex thread

Day three was the closing day of NVIDIA's GTC Taipei presence inside Computex. Day one had unveiled the RTX Spark Superchip for PCs, the announcement that drove the consumer-side Computex news cycle and pulled the rug from under Intel and Qualcomm in the AI-PC silicon segment. Day two extended the AI-PC story with partner-card and partner-laptop announcements from Dell and Lenovo. Day three's TSMC partnership and AI for Media announcements close out the conference for NVIDIA. The cumulative read is that NVIDIA's strategy has moved past the question of whether it dominates the datacenter and toward the question of how broad the AI accelerator's customer base ultimately becomes. The TSMC partnership is the latest data point that the customer base extends to the upstream silicon supplier.

Application

ChatGPT for Excel and Google Sheets opens globally, the in-cell sidebar moves out of preview

OpenAI Help Center · 2026-06-03 · Source ↗ ·

ChatGPT for Excel and Google Sheets is now available globally. The product brings ChatGPT into a sidebar in both spreadsheet environments for trackers, budgets, formulas, multi-tab files, scenario work and cleanup tasks. Business customers had a free preview through June 2; usage now follows plan credits. Enterprise, Edu and K-12 customers move to the same model on the same date. The integration supports Skills and apps where available.

OpenAI's spreadsheet sidebar reached global general availability across both Microsoft Excel and Google Sheets on June 3. The product runs as an in-app sidebar that lets users build, update and understand spreadsheets without leaving the workbook. The supported task surface includes trackers, budgets, formulas, multi-tab files, scenario work and spreadsheet cleanup. The integration also surfaces Skills and apps where available, which are the OpenAI-native interfaces to third-party tools and data sources.

The pricing transition

The free-preview window through June 2 ran for Business, Enterprise, Edu and K-12 customers. After June 2, usage shifts to the standard plan-credits model that governs the rest of the ChatGPT business product line. The transition is the actual June 3 news: the global availability matters operationally because the metering changed at the same moment, and customers running the product as a free preview now have to decide whether to keep the spend or pull back. OpenAI's pricing posture is that the in-spreadsheet sidebar is a metered ChatGPT usage surface, not a separately-priced product, which simplifies the procurement story for the enterprise customers who already license ChatGPT Business.

The competitive frame

The Excel-and-Sheets sidebar lands against Microsoft 365 Copilot's Excel integration, against Google's native Workspace AI, and against Anthropic's Claude file-upload workflow. OpenAI's distinctive position is that the sidebar sits inside the user's existing spreadsheet environment regardless of which vendor that environment comes from, which is a structurally easier procurement story for customers running mixed Excel-and-Sheets deployments. The Microsoft Copilot pitch is tighter integration with the rest of the Office stack; the OpenAI pitch is vendor neutrality. The actual usage telemetry over the next two quarters will determine which positioning wins the enterprise wallet.

Tech

Snowflake Datastream lands as managed Kafka, real-time streams meet AI agents inside the perimeter

Snowflake Newsroom / iTWire · 2026-06-03 · Source ↗ ·

Snowflake unveiled Datastream as a managed streaming service for Apache Kafka, in private preview at announcement. The service combines real-time ingestion with the Cortex AI pipeline in a single billed plane, letting customers wire a Kafka topic into a Cortex agent without operating a separate Kafka cluster and without exporting data outside the Snowflake compliance boundary. The Cortex Sense context layer ties the streaming data to business definitions for agent consumption.

Snowflake Datastream is the fourth of the Snowflake Summit 2026 launches that matter operationally. The service is a managed streaming offering for Apache Kafka, sized to handle the real-time ingestion workloads that enterprise customers have historically run on Confluent or on a self-managed Kafka cluster. The service is in private preview at launch and is structured as a Snowflake-billed managed offering rather than as a partner integration.

The use case is the streaming-to-agent workflow that has been the hardest piece of enterprise AI to operationalize. The conventional pattern involves a Kafka cluster operated by the customer or by Confluent, a transformation layer in dbt or Flink, an export to the warehouse, and finally an agent invocation. Each step in the pattern is a separate operational surface with its own SLA, billing and security review. Datastream collapses the pattern into a single managed plane within Snowflake's perimeter, with the customer's existing Snowflake identity, audit and data-residency controls already covering the new surface.

The Cortex Sense connection

Datastream pairs with Cortex Sense, the context layer Snowflake also introduced this week, which brings together data, business definitions and operational knowledge for AI agents. The combination is the load-bearing piece: a streaming topic is only useful to an agent if the agent has the business context to interpret what the stream means. Cortex Sense supplies that context as a queryable layer over the customer's metadata, with agents pulling business definitions, lineage and policy from Sense at runtime rather than having those definitions embedded in the agent's prompt.

The competitive frame is Confluent's Tableflow and Databricks's Delta Live Tables. Snowflake's pitch is governance: the same audit, identity and data-residency posture customers already run for their warehouse extends to the streaming surface without a separate review. Pricing follows the Snowflake credit model against stream throughput and storage. Datastream is expected to reach general availability over the second half of 2026.

Tech

Snowflake Cortex Training opens managed-GPU fine-tuning of Qwen and Mistral inside the Snowflake boundary

Snowflake Newsroom · 2026-06-03 · Source ↗ ·

Cortex Training provides managed-GPU infrastructure that lets organizations fine-tune open-source models like Qwen and Mistral against proprietary Snowflake-resident data without exporting any training data outside the perimeter. The service is targeted at regulated industries that have to keep customer and patient data inside the warehouse compliance boundary, and at customers that have wanted to graduate from prompt engineering on hosted models to actual weight-level customization on their own corpora.

Cortex Training is Snowflake's answer to the enterprise question of how a regulated customer fine-tunes an open-source model against its own data without the legal, security and audit headaches of moving the training data to a third-party vendor. The service provides managed-GPU infrastructure inside Snowflake's compliance perimeter, exposes Qwen and Mistral as the launch list of supported base models, and lets the customer run a supervised fine-tuning workflow against data that never leaves the Snowflake warehouse.

The launch list is small by design: Qwen and Mistral are the two open-weights families with both broad enterprise adoption and licensing structures clean enough to support a managed-fine-tuning offering. The service catalog is expected to expand as additional open-weights base models, including Llama, Gemma, OLMo and DeepSeek, clear Snowflake's security and licensing review. The business pitch is the regulated-industry use case: financial services, healthcare and government customers cannot move training data to AWS Bedrock or to a separate fine-tuning vendor, and Cortex Training collapses the entire workflow inside Snowflake.

The economic and technical structure

Pricing is metered against GPU hours, with a separate base-model licensing add-on. The architecture is Snowflake-managed: the customer never sees the underlying GPU cluster, and the fine-tuned model artifacts are persisted inside the Snowflake account boundary as governed objects. The trained model can be served from Cortex's existing inference endpoint or exported to the customer's own deployment target, with the export gated by the customer's data-classification policy.

The technical bet is that the next round of enterprise AI competition is fought at the data-customization tier rather than at the base-model tier. The frontier models from Anthropic, OpenAI and Google are converging on similar capability profiles; the durable enterprise differentiator is the model's behavior on the customer's own data and the institution-specific reasoning the model can produce. Snowflake's argument is that the fine-tuning workstream belongs inside the data plane, not bolted onto it from the outside.

ApplicationTech

Microsoft opens a Discovery desktop app preview for individual researchers and small teams

Microsoft Azure Blog · 2026-06-03 · Source ↗ ·

The Microsoft Discovery desktop app entered private preview for researchers, students and small scientific teams on June 3. The local experience exposes the same agentic R&D loop Discovery runs in Azure Foundry, scoped to single-researcher and small-team workflows that don't require the full enterprise deployment. The app is the on-ramp for academic and individual users who want to use Discovery without standing up a corporate Foundry tenancy.

Alongside the Microsoft Discovery general-availability announcement, Microsoft opened a desktop-app preview targeted at the audience the enterprise Discovery deployment doesn't reach: individual researchers, graduate students, and small scientific teams. The app is a local Windows-and-macOS experience that exposes the same evidence-to-hypothesis loop, the same agent model and the same connector library the enterprise Discovery deployment runs in Azure Foundry, but scoped to the access patterns that single-researcher and small-team workflows actually have.

The pitch is that academic and individual users have historically been the audience that propagates new research-tooling primitives into the broader scientific community. By giving that audience a frictionless on-ramp, Microsoft is betting that Discovery's institutional adoption will follow rather than precede the bottoms-up adoption by the researchers who set the experimental-tooling defaults at their institutions. The preview is gated through the Microsoft Foundry portal and requires a Microsoft account; pricing for the GA version of the desktop app has not been disclosed.

What the desktop preview ships with

The preview ships with the Microsoft Discovery Engine running locally for state management and the agent layer connected to Azure Foundry-hosted base models for the actual reasoning work. The connector library at preview covers the standard scientific tooling surface: arXiv, PubMed, Reaxys for chemistry, the AlphaFold Protein Structure Database, Materials Project for materials science, and a default set of HPC simulation tools that can be invoked through the agent layer. Microsoft has said the connector library will grow through the preview window based on the workflows researchers actually run.

The desktop app is consumed via Microsoft account sign-in; researchers running on an institutional Azure Foundry deployment can switch the app to use their tenant's resources rather than the Microsoft-managed preview backend. The pricing path for the GA version is expected to converge on Microsoft 365's consumer and academic pricing rather than on Azure Foundry's enterprise consumption-billed pricing.

Tech

Apache Iceberg v3 reaches general availability through Snowflake Horizon Catalog with bi-directional engine access

Constellation Research / Snowflake · 2026-06-03 · Source ↗ ·

Snowflake announced general availability of Apache Iceberg v3 on June 3, with the broadest feature support of any commercial engine for the open table format. Horizon Catalog, now powered by Apache Polaris, exposes Snowflake-managed Iceberg tables for bi-directional read and write through external engines including Databricks, Trino, BigQuery and DuckDB using open standards. Iceberg v3 is the version that adds row-lineage, deletion vectors and column-level encryption.

Apache Iceberg v3 is the table format the lakehouse community has been waiting for since the v2 release in 2023. The v3 spec adds row-level lineage, deletion vectors for write-efficient deletes, and column-level encryption for fine-grained access control. Snowflake's June 3 announcement is that v3 is now the default format for Snowflake-managed Iceberg tables, with the broadest feature support of any commercial engine that backs Iceberg. The Snowflake-managed Iceberg storage tier is the production deployment surface.

The largest piece of the announcement is the Horizon Catalog interoperability. Horizon, the metadata layer Snowflake has been building around Apache Polaris, now exposes Snowflake-managed Iceberg tables for bi-directional read-and-write access through external engines, using open standards-based access controls. The practical implication is that a Snowflake customer's Iceberg tables can be queried and modified by Databricks, Trino, Presto, BigQuery, DuckDB or any other Polaris-compatible engine, with the access control enforced uniformly through the Polaris layer.

The bet on engine neutrality

The strategic frame for the Iceberg v3 GA is that the table-format wars have ended and the engine wars are starting. Snowflake's Iceberg-v3 position concedes the format question to the open standard and competes on the runtime, the catalog, the governance posture and the surrounding services. The Polaris-powered Horizon Catalog is the structural piece of that strategy: by making the metadata layer open and neutral, Snowflake positions itself as the highest-quality runtime against a shared metadata standard rather than as the proprietary format owner trying to lock customers in.

The competitive frame is Databricks's Unity Catalog. Databricks's pitch is that Unity is the single source of truth for governance across the lakehouse and that Databricks-flavored Delta Lake is the better-engineered table format. Snowflake's pitch with Iceberg v3 GA and Polaris-powered Horizon is that openness wins over time and that the customer with multi-engine environments will choose the open-standard combination over a Databricks-controlled equivalent. The Data + AI Summit two weeks from now will be Databricks's chance to respond.

ApplicationTech

Nvidia previews AI for Media at GTC Taipei, real-time blocks for live production, localization and synthetic-video detection

NVIDIA GTC Taipei · 2026-06-03 · Source ↗ ·

NVIDIA previewed a new AI for Media technology family at GTC Taipei day three. The building blocks cover real-time live production, localization, content analysis and synthetic-video detection, aimed at broadcasters, streaming platforms and developers. The synthetic-video-detection piece is the strategically loaded line: NVIDIA is positioning its media stack as the verification substrate newsrooms and platforms increasingly need as generative-video tools become widely available.

NVIDIA's AI for Media announcement at GTC Taipei is part of the broader pivot the company has been making from accelerator vendor to full-stack AI platform supplier. The new technology family targets broadcasters, streaming platforms and the developers building tools for those audiences. The four named building blocks are real-time live production, localization, content analysis and synthetic-video detection.

Real-time live production is the workflow where the latency budget is in the tens of milliseconds: live sports, live news, live event coverage. NVIDIA's pitch is that the GPU-accelerated AI inference stack collapses the time it takes to do production-quality processing, switching, graphics overlays and quality control into the live event itself, removing the post-production stage from the workflow for the live-broadcast use case. Localization is the AI-driven translation and dubbing workstream that has been transforming how international content is distributed; NVIDIA's offering is positioned against the existing third-party localization vendors as a more deeply integrated alternative.

Synthetic-video detection

The synthetic-video-detection piece is the most strategically loaded. Newsrooms, streaming platforms and political-advertising marketplaces are facing the verification problem head-on as generative-video tools become widely available. NVIDIA's positioning is that the media stack should ship verification as a first-class capability alongside production and localization, rather than treating it as an afterthought. The technical approach is multi-signal: cryptographic provenance where it exists, model-fingerprint detection against known generative-video models, and behavioral analysis of the video's frame-to-frame consistency. The announcement did not name the broadcaster customers for the AI for Media preview, but Reuters, Sky and Comcast were named in the broader NVIDIA Media partner ecosystem at GTC Taipei.

The announcement also covered DLSS 4.5 Ray Reconstruction on the consumer-graphics side and the expansion of RTX Spark availability to additional partner OEMs. The consumer-graphics updates are tangential to the AI-for-Media story but underline the breadth of NVIDIA's June 3 GTC Taipei keynote, which spanned data-center, semiconductor-fab, professional-media and consumer-graphics audiences in a single morning.

Safety

Anthropic publishes machine-readable MITRE ATT&CK mapping for AI-orchestrated attack chains, framework gap flagged

Anthropic Threat Intelligence · 2026-06-03 · Source ↗ ·

Anthropic released the full per-technique mapping behind the year-long cyber-threat analysis as a machine-readable contribution to the MITRE ATT&CK community. The release identifies a gap in the existing framework's vocabulary: ATT&CK has no first-class terminology for agentic orchestration or for the autonomous chaining of attack stages. Anthropic is in active discussion with MITRE on extending the framework to capture these AI-enabled patterns and is publishing the underlying dataset to seed that workstream.

Alongside the headline report, Anthropic released the full per-technique mapping from the 832-account analysis as a machine-readable contribution to the MITRE ATT&CK community. The dataset gives security vendors, CIRTs and academic researchers a structured view of which MITRE techniques are most heavily AI-assisted and how the per-technique pattern shifts between low-skill and high-skill actor populations.

The technical contribution comes with a framework-gap observation. MITRE ATT&CK was built around the deterministic attacker tradecraft that defined the late-2010s threat landscape. The framework has rich vocabulary for individual technique categories but treats the chaining of techniques as a sequential, attacker-driven process. The cases Anthropic surfaced, especially the November 2025 espionage operation that used 30 techniques across 13 tactics under autonomous orchestration, do not fit the sequential, attacker-driven model. The chaining is autonomous; the planning is in the LLM; the human attacker's role is the initial mission specification and the post-execution review.

The proposed framework extension

Anthropic's recommendation to MITRE is that ATT&CK add first-class terminology for three patterns. The first is agentic orchestration as a tactic in its own right, distinct from the existing tactics that describe individual technique categories. The second is autonomous chaining, which would capture the cases where the agent decides on the next technique to deploy based on the result of the previous step rather than executing a pre-planned sequence. The third is what Anthropic calls operational tempo, which would capture the rate at which an AI-orchestrated attack can iterate compared to a human-orchestrated attack and which has become a security-relevant property in its own right.

MITRE has not committed publicly to incorporating the extensions. The Anthropic publication is the opening of the conversation rather than the resolution, and the timeline for any framework update is measured in quarters rather than weeks. In the interim, the published mapping is usable directly by security vendors building detection rules against the AI-orchestrated tradecraft pattern.

Why the publication matters

The publication is the second consequential Anthropic-led contribution to the security community in 2026, after the November 2025 disclosure of the first AI-orchestrated cyber-espionage operation. The pattern Anthropic is establishing is that the frontier-model developer takes responsibility for surfacing the misuse data that the security community can act on. The pattern is consistent with the Verizon Data Breach Investigations Report collaboration covered in the broader Anthropic story and signals that Anthropic intends to be a load-bearing participant in the cybersecurity ecosystem rather than just a tool vendor that ships safeguards.

Tech

Snowflake Cortex Sense ships as the business-context layer agents query at runtime

Snowflake Newsroom · 2026-06-03 · Source ↗ ·

Cortex Sense is Snowflake's new context layer that brings together data, business definitions and operational knowledge for consumption by AI agents at runtime. The premise is that agents need business context inline with the data, not embedded in their prompts. Sense exposes metadata, lineage, ownership and policy as a queryable layer agents can call during execution, so the same agent can run safely against a different customer's data without re-prompting.

Cortex Sense is the piece of the Snowflake Summit announcement that is the least flashy but the most architecturally consequential. Sense is the runtime context layer agents query to understand what the data they are operating on actually means. The premise is that an agent's behavior is only correct if the agent has the business context to interpret the data correctly. The conventional pattern embeds that context in the agent's prompt, which couples the agent to a single customer's data model and forces every change in the customer's metadata to be re-stitched into the prompt.

Sense factors the context out of the prompt and into a queryable layer that agents call at runtime. The layer exposes the customer's metadata, the lineage of any given data object, the ownership and policy assigned to that object, and the business definitions that describe what the object means in the customer's terms. The agent then constructs its action by combining the customer's Sense layer with the same agent code that runs against every other customer's Sense layer, which makes the agent portable across customer deployments in a way the prompt-embedded-context pattern does not.

The MCP comparison

The closest external analog is the Anthropic-led Model Context Protocol that has become the de facto standard for agent-tool integration over the past year. Sense is to data what MCP is to tools: a standardized interface the agent calls at runtime rather than a per-deployment configuration. The Snowflake-specific differentiation is that Sense is tied to the customer's data plane and to the warehouse-level governance posture, which makes the context layer audit-able through the same primitives that govern the data itself.

Sense is in private preview at the Snowflake Summit announcement and is expected to reach general availability over the second half of 2026 alongside Datastream. The combined positioning is that the agent layer should be portable, but the data layer and the context layer should travel together inside the customer's compliance boundary.