Microsoft Build 2026 & More (0603)

1. Microsoft Build 2026 — Full Unveiling of Proprietary AI Models and Agent Platform

• Core Source

“Unveiled proprietary model family MAI. Includes reasoning model MAI-Thinking-1, image generation MAI-Image-2.5, and voice, transcription, and coding models”

“MAI-Thinking-1 supports 35 billion active parameters and 256K context. Claims preference advantage over Sonnet 4.6 and coding performance on par with Opus 4.6”

“Majorana 2 unveiled. Average qubit coherence time exceeding 20 seconds, reliability improved 1,000x over previous generation. Target of building a scalable quantum computer by 2029”

“AI agent platform strategy in full swing. GitHub, Foundry, Copilot, Windows, and Azure to be integrated into a single agent stack”

“Windows transitioning to an agent-native platform. OS-level sandbox and security isolation provided through MXC (Microsoft Execution Containers)”

• Expected Impact

The core of what Microsoft revealed at Build 2026 can be broken down into three areas.

First, the unveiling of the proprietary AI model lineup (MAI) aimed at reducing dependence on OpenAI. MAI-Thinking-1 supports 35 billion active parameters and 256K context, with Microsoft claiming a preference advantage over Anthropic’s Sonnet 4.6 and coding performance on par with Opus 4.6. Microsoft had previously relied entirely on OpenAI models through Azure, making the development of a proprietary model portfolio a structurally significant shift in terms of improving the cost structure of AI services and strengthening negotiating leverage with third-party model providers.

Second, a full integration strategy for the AI agent platform. The declaration to unify GitHub, Foundry, Copilot, Windows, and Azure into a single agent stack signals Microsoft’s intent to reposition itself from a cloud and SaaS provider into a platform operator commanding the entire enterprise AI workflow. Products such as Microsoft Scout (an autonomous work agent based on Teams and Outlook), Frontier Tuning (reinforcement learning using internal enterprise data), Agent 365 and MDASH (agent governance and security), and GitHub Copilot App (integrated parallel agent development, code review, CI, and merging) are all product lines designed to lower the barrier to AI adoption for enterprise customers while reinforcing Microsoft ecosystem lock-in. Transitioning Windows to an agent-native platform and providing OS-level sandbox isolation via MXC is a strategy to bring agent execution environments onto Windows while meeting enterprise security requirements.

Third, an expansion of hardware strategy. The Surface RTX Spark Dev Box (up to 1 PFLOPS of AI performance, 128GB unified memory, local execution of models up to 120B parameters) is a compact developer PC from Microsoft running on NVIDIA’s RTX Spark superchip, illustrating the direction of offloading some AI inference workloads from cloud to edge devices. Additionally, the Microsoft CEO’s preview of the Cobalt 200, a proprietary AI server processor, signals that the strategy of reducing cloud compute costs through custom silicon has entered full execution mode.

The unveiling of Majorana 2 is also noteworthy on the quantum computing front. With an average qubit coherence time exceeding 20 seconds and reliability improved 1,000x over the previous generation, Microsoft has set a target of building a scalable quantum computer by 2029.

In sum, Build 2026 may mark a pivotal moment in which Microsoft accelerates its transition from dependence on OpenAI models to becoming an operator of its own model, platform, and hardware vertical. If proprietary model performance reaches a level that can genuinely compete with external models, Microsoft’s AI-related profitability structure has room for improvement over the medium to long term.

2. OpenAI, Intelligence at Work — Full-Scale Expansion into Knowledge Work

• Core Source

“Codex WAU surpasses 5 million. Growth of more than 6x since desktop app launch in February. Knowledge workers now account for approximately 20% of total users”

“Codex adoption among knowledge workers is more than 3x faster than among developers. Individual users are also growing more than 4x faster than developers”

“The fastest-growing tasks among knowledge workers are data analysis (+110% WoW), research (+37%), and knowledge output creation (+36%)”

“The defining behavioral shift is parallel tasking. Approximately 50% of users now run multiple Codex tasks simultaneously, up significantly from less than 33% in mid-April”

“Enterprise customer count surpasses 2 million, doubling year-over-year. Enterprise revenue accounts for approximately 40% of total revenue, with a target of reaching 50% by year-end”

• Expected Impact

The central message from OpenAI’s Intelligence at Work event is that Codex is rapidly expanding from a developer tool into an AI platform capable of replacing knowledge work across the board.

First, the pace of growth is exceptional. Codex WAU has surpassed 5 million, growing more than 6x since the desktop app launched in February, and approximately 400% since the start of the year. What is particularly notable is that the growth is not being driven by developers. Non-developer knowledge workers are adopting Codex more than 3x faster than developers, and individual users are growing more than 4x faster. This reflects Codex’s market expanding beyond the developer tools market to encompass the entire knowledge work market, which accounts for approximately 40% of U.S. workers, or 72 million people.

Second, the behavioral shift among users is structural. Approximately 50% of users now run multiple Codex tasks in parallel simultaneously, up sharply from less than 33% in mid-April. OpenAI interprets this as a signal that users are transitioning from individual task performers to orchestrators managing multiple AI workflows. The fastest-growing use cases are data analysis (+110% WoW), research (+37%), and knowledge output creation (+36%), demonstrating that penetration is accelerating across core knowledge work, well beyond simple code generation.

Third, a transition in enterprise revenue structure is underway. Enterprise customer count has surpassed 2 million, doubling year-over-year, with enterprise revenue accounting for approximately 40% of total revenue and projected to reach 50% by year-end. The launch of six role-specific plugins (data analysis, creative production, sales, product design, public equity investing, and investment banking), along with planned additions in corporate finance, law, and marketing, signals the full-scale adoption of a vertical SaaS strategy that goes deep into specific job functions. Furthermore, Codex is set to be integrated into ChatGPT within the next few weeks, and with ChatGPT WAU approaching approximately 1 billion, this integration could dramatically expand Codex’s potential reach.

Fourth, improvements in model efficiency are also an important signal. GPT-5.5 can reportedly generate comparable outputs using approximately one-third the token consumption of its predecessor, and the model release cycle has shortened from an average of 15 months to approximately 6 weeks. As token efficiency improves, OpenAI’s compute cost burden eases and a structure emerges in which more customers can be acquired at lower prices, making this a factor that simultaneously supports profitability and market expansion.

3. Anthropic, Project Glasswing Expansion — AI Cybersecurity Moves into Critical Infrastructure

• Core Source

“Project Glasswing, a cybersecurity program based on Claude Mythos Preview, expanded from 50 participating organizations to approximately 200”

“Early participating organizations have collectively identified more than 10,000 high-severity and critical vulnerabilities to date”

“Multiple AI companies are projected to acquire Mythos-level cyber capabilities within the next 6 to 12 months. The need to advance defensive systems is emphasized”

“The core bottleneck in cybersecurity is assessed to be not the discovery of vulnerabilities, but the process of verification, disclosure, and patching”

“Anthropic’s latest AI model, Claude Mythos, has been available only to select partners since early April, based on the judgment that it could give rise to potential cyberthreats.”

• Expected Impact

The core of this announcement is twofold: that Claude Mythos is producing tangible results in the cybersecurity domain, and that Anthropic is rapidly establishing a first-mover position in the AI security market.

First, Mythos’s cybersecurity capabilities have been validated in the field. The 50 early participating organizations in Project Glasswing have collectively identified more than 10,000 high-severity and critical vulnerabilities to date. Anthropic assessed that the core bottleneck in cybersecurity lies not in vulnerability discovery but in the verification, disclosure, and patching process, and noted that Mythos is being used not only for vulnerability detection but also for patch writing, penetration testing, and threat detection automation. This demonstrates that Mythos is establishing itself as a tool covering the entire security workflow. Anthropic also unveiled Claude Security, a code scanning and patch suggestion solution based on Claude Opus 4.8.

Second, the pace and scope of the expansion in participation is notable. The number of participating organizations has quadrupled from 50 to approximately 200, with new participants including operators of critical infrastructure such as power, water, healthcare, and communications across more than 15 countries, as well as key open-source maintenance organizations. At launch in early April, Mythos was restricted to select partners due to the potential for cyberthreats, but following approximately two months of developing countermeasures, access has been expanded to more than 150 partners across more than 15 countries. AI security platforms deployed to critical infrastructure operators tend to have high switching costs once adopted, making Anthropic’s early accumulation of references in this market highly significant over the medium to long term.

Third, Anthropic’s view of the evolving competitive landscape is also worth noting. Anthropic projects that multiple AI companies will acquire Mythos-level cyber capabilities within the next 6 to 12 months, emphasizing the need to advance defensive systems. This reflects an awareness within the industry that the proliferation of AI-based cyberattack capabilities is imminent, while also signaling Anthropic’s strategic intent to establish a dominant position on the defensive side of the ecosystem. The explicit statement that the goal of Project Glasswing extends beyond vulnerability detection to supporting a transformation in how cybersecurity operations are conducted across the software industry as a whole underscores this intent.

4. Palo Alto Networks FY3Q26 Earnings — AI Security Platform Reshaping Competitive Landscape

• Core Source

“Next-Generation Security (NGS) ARR reached $8.13 billion, up 60% year-over-year, while revenue grew 31% to $3.0 billion. Organic NGS ARR, excluding acquisition effects, also grew 28% to $6.5 billion, confirming strong demand across all regions.”

“RPO (Remaining Performance Obligations) grew 36% year-over-year to $18.4 billion, with Current RPO also growing 34% to $8.3 billion. Hardware backlog and next-generation firewall (NGFW) orders also rose approximately 40%, strengthening future revenue visibility.”

“Prisma AIRS customer count surpassed 300, with ARR of $100 million projected within the next few quarters. XSIAM ARR exceeded $600 million, growing 100% year-over-year. Chronosphere ARR surpassed $300 million, with adoption expanding among AI-native customers.”

“Integration of CyberArk and Chronosphere is progressing faster than initially expected. More than 1,000 cross-selling opportunities have emerged, and management noted that the timeline for reaching profitability targets could be accelerated by 3 to 6 months.”

“Adjusted free cash flow (FCF) for Q3 was $910 million. Trailing twelve-month adjusted FCF reached $4.08 billion, achieving an FCF margin of 38.5%.”

• Expected Impact

The key takeaway from Palo Alto Networks’ FY3Q26 results is that the AI security platform strategy is beginning to be validated in numbers.

First, the quality of growth is high. Revenue grew 31% year-over-year to $3.0 billion, beating consensus, while NGS ARR grew 60% to $8.1 billion. Particularly noteworthy is that organic NGS ARR, excluding acquisition effects, also grew 28% to $6.5 billion. This demonstrates that the underlying strength of the core business is intact, and that top-line growth is not solely dependent on M&A. RPO also grew 36% to $18.4 billion, elevating future revenue visibility.

Second, adoption of the AI security platform product suite is accelerating. Prisma AIRS surpassed 300 customers, XSIAM ARR grew 100% year-over-year to $600 million, and Chronosphere ARR expanded to more than $300 million. These three products cover AI threat detection, AI-driven security operations (SOC) automation, and observability, respectively. As AI infrastructure expands, so too does the attack surface, and Palo Alto Networks is building a position that intensively absorbs the defensive demand responding to increasingly sophisticated AI-based attacks.

Third, acquisition integration synergies are materializing faster than expected. The CyberArk and Chronosphere acquisitions have generated more than 1,000 cross-selling opportunities, and management’s comment that the timeline for reaching profitability targets could be pulled forward by 3 to 6 months reinforces confidence in execution capability. FY26 annual NGS ARR guidance was raised to $8.9 billion to $8.95 billion, and revenue guidance was also raised to $11.415 billion to $11.425 billion.

However, risk factors also exist. Hardware prices were raised approximately 10% due to rising memory and storage costs, and stock-based compensation (SBC) related to M&A rose to approximately 17% of revenue, leaving the company still reporting a net loss on a GAAP basis (-$0.22 per share). The growing share of large contracts — including a $200 million AI research lab contract and an $80 million utility company contract — means that execution risk is expanding alongside contract size, which warrants ongoing monitoring.

5. SK Hynix, DRAM Production Capacity to Double by 2030

• Core Source

“SK Group Chairman Chey Tae-won presented a plan to double SK Hynix’s DRAM WSPM over the next five years.”

“SK Hynix held 530,000 DRAM WSPM as of end-2025, with 350,000 in Korea and 180,000 in China. We expect 745,000 by end-2028 and 865,000 by end-2029.”

“In JPMorgan’s base case scenario, the supply shortage in 2027 will be worse than in 2026, and the absolute supply shortage is expected to persist through 2028.”

“In our view, more than 50 EUV tools will be needed in the 2028–2030 period, and SK Hynix’s official infrastructure completion target is December 2030.”

“SK Hynix Chairman Chey’s comment that memory supply bottlenecks will persist through 2030 was also seen as a positive.”

• Expected Impact

Chairman Chey Tae-won’s announcement is noteworthy not merely as a disclosure of investment plans, but as a direct confirmation of the top leadership of a major memory supplier’s outlook on industry conditions. JPMorgan cited and evaluated the chairman’s remarks directly.

First, the scale and timeline of the expansion plan are concrete. SK Hynix’s DRAM WSPM stood at 530,000 wafers as of end-2025. JPMorgan projects a staged expansion to 745,000 by end-2028 and 865,000 by end-2029, with capacity exceeding 1 million wafers once the first phase of the Yongin cluster is fully occupied by end-2030. This represents a plan to approximately double production capacity over the next five years. However, JPMorgan noted that the realization of this plan is contingent on resolving variables including the procurement of more than 50 EUV tools and the construction of EPC infrastructure in the 2028–2030 period.

Second, JPMorgan interprets this expansion as a response to supply shortages rather than a peak-out signal. Historically, announcements of large-scale capacity expansions by memory suppliers have often been received negatively by markets. However, JPMorgan expects the supply shortage to worsen in 2027 compared to 2026, with the absolute supply shortage persisting through 2028, and therefore characterizes this expansion as an unavoidable response to keep pace with demand. Chairman Chey’s direct statement that memory supply bottlenecks will persist through 2030 is consistent with JPMorgan’s view.

6. TrendForce — HBM: Profitability Reversal in 2026, Price Surge Structure Taking Shape for 2027

• Core Source

“Starting in Q1 2026, HBM’s per-wafer revenue has fallen below that of DDR5 64GB RDIMMs. As a result, suppliers are rebalancing production between HBM and conventional DRAM to protect profitability.”

“The ‘crowding-out effect,’ whereby larger die sizes and growing demand erode conventional DRAM production capacity, is expected to intensify further in 2027. This provides suppliers with strong grounds to push for significant price increases in 2027 HBM contract negotiations.”

“HBM capacity per AI ASIC increasing from 96GB–192GB previously to 216GB–288GB”

“NVIDIA Rubin Ultra is expected to expand HBM capacity per GPU to 384GB”

“HBM wafer input share forecast: end-2025: 18%, end-2026: 22%, end-2027: 30%”

• Expected Impact

The core of TrendForce’s analysis is that both the incentive for suppliers to raise prices and demand-side pressure are simultaneously intensifying in the HBM market.

First, the profitability reversal is strengthening suppliers’ incentive to raise prices. Since Q1 2026, HBM’s per-wafer revenue has fallen below that of DDR5 64GB RDIMMs. This is the result of HBM’s annual contract structure failing to immediately capture the surge in conventional DRAM prices that began in the second half of 2025. With HBM profitability now lower than that of conventional DRAM, Samsung, SK Hynix, and Micron have strong incentives to push for significant price increases in HBM4 contract negotiations for 2027. The profitability reversal serves as a structural basis that suppliers can use to justify price hikes in 2027.

Second, on the demand side, HBM content per chip is rising sharply with each generation. In 2026, AI ASICs are the primary driver of demand growth, with HBM capacity per AI chip increasing from 96GB–192GB to 216GB–288GB. In 2027, NVIDIA’s Rubin Ultra platform is expected to expand HBM capacity per GPU to 384GB, and growing deployment volumes of AI ASIC platforms such as Google TPUs are expected to further accelerate HBM bit demand.

The outcome of these converging pressures is reflected directly in the production share outlook. HBM’s share of total DRAM wafer inputs is projected to expand from 18% at end-2025 to 22% at end-2026 and 30% at end-2027, while HBM’s share of total DRAM bit supply is expected to grow from 8% to 13% over the same period. As HBM’s production share rises, conventional DRAM supply is correspondingly eroded, creating upward price pressure on conventional DRAM as well. As a result, in an environment where demand pressure structurally exceeds supply, suppliers are positioned to push through significant price increases in 2027 HBM4 contract negotiations. The favorable supply-demand environment for leading HBM suppliers such as SK Hynix and Samsung is likely to persist through 2027.

7. Marvell × NVIDIA, Optical Interconnect Comes of Age — ‘Copper Wall’ Declared

• Core Source

“Compute (GPU) → Memory (HBM) → Connectivity: the next bottleneck is how fast data can be moved”

“World’s first 102.4Tbps AI-dedicated switch launched. 51.2T CPO switch (16×3.2T optical engines) demonstrated live on stage. 1.6Tbps 2nm coherent optical solution — sampling expected within the year (world’s first)”

“Use copper where you can, optical where you can’t” — Jensen Huang

“NVIDIA’s $2 billion investment in Marvell reaffirmed. Collaboration extended across optical, silicon photonics, and NVLink Fusion”

• Expected Impact

The central message jointly delivered by Marvell and NVIDIA at Computex 2026 is a declaration that the next bottleneck in AI infrastructure has shifted to Connectivity.

First, the shift in bottleneck creates a new zone of beneficiaries. Marvell CEO Matt Murphy presented AI infrastructure development in three stages: Compute (GPU) → Memory (HBM) → Connectivity. If NVIDIA and SK Hynix were the representative beneficiaries of the first two stages dominated by GPUs and HBM, then companies with optical interconnect technology stand to receive structural benefits in the connectivity stage. Marvell stated that it has been preparing for this transition for years, and NVIDIA confirmed as much by reaffirming its $2 billion investment in Marvell and expanding the scope of collaboration to cover optical technology, silicon photonics, and NVLink Fusion broadly.

Second, the ‘Copper Wall’ declaration formalizes the inevitability of the optical transition. Jensen Huang drew a clear line: “Use copper where you can, optical where you can’t.” As bandwidth requirements escalate in rack-scale connectivity, copper runs into physical limitations, making the transition to optical technology unavoidable. In response, Marvell launched the world’s first 102.4Tbps AI-dedicated switch and announced the sampling of the world’s first 1.6Tbps 2nm coherent optical solution within the year. The solution architecture is structured by distance, with Marvell COLORZ 1600 (4th-generation silicon photonics) applied to inter-datacenter links (hundreds to thousands of kilometers), and PAM4-based power-optimized optical links applied to intra-datacenter connections (up to ~500 meters).

Jensen Huang singled out Marvell as the “next trillion-dollar company,” emphasizing that connectivity is the truly indispensable element in datacenters where computing workloads are distributed across thousands of chips.