Microsoft Build 2026 announced an end-to-end agentic AI stack. COMPUTEX Taipei confirmed heterogeneous AI infrastructure across ARM, Marvell, Intel, Qualcomm, and NVIDIA. Alphabet raised $80 billion. Cisco Live repositioned the network as the AI platform. Patrick Moorhead and Daniel Newman break it all down alongside earnings from Broadcom, HPE, Palo Alto Networks, and CrowdStrike, plus the token cost conversation, the edge AI push, and what Palantir and Oracle are saying about proprietary data as the real AI moat. The handpicked topics for this week are: Microsoft Build 2026 Announced an End-to-End Agentic AI Stack: Microsoft shipped MAI-Thinking-1, its first homegrown thinking model, alongside Scout, Microsoft IQ, Project Solara, and a Majorana 2 quantum update targeting a 2029 commercial timeline with claims of a 1,000x reliability gain. Pat describes MAI-Thinking-1 as likely better than Sonnet 4.6 in blind testing and delivering close to GPT 5.5 quality at a far lower cost. Scout is Microsoft's first autopilot agent, anchoring the M365 Agent Suite with Office Pilot Agent Mode and Agent 365. Microsoft IQ serves as the context layer, integrating M365, business data, boundary IQ, and web IQ with GitHub Copilot, Foundry, and Copilot Studio. Project Solara is a new Android-based platform built for agent-first devices across transportation, retail, and hospital settings. Microsoft also added 83 Unix commands to the Windows stack. Dan frames Microsoft's real play as distribution, not frontier model development, noting that the open model ecosystem being pulled into the platform will matter more to CFOs managing token costs at scale. (The Decode) The AI Stack Goes Multi-Silicon — COMPUTEX Taipei 2026 Confirms Heterogeneous AI Infrastructure: ARM's AGI CPU is in production with Google moving its TPU head node to ARM, and adding Oracle and ByteDance as new customers. ARM also introduced a new switch, the TT100, and put the 51T CPO switch on stage. Marvell received a trillion-dollar company endorsement from Jensen Huang, adding $90 billion in market cap on the comment alone. Intel announced disaggregated inference details and Xeon 6+ Clearwater Forest, its first 18A data center processor. Vista Equity and Cambium Capital announced a NeoCloud called Vector Core Compute, with Xeon 6 handling orchestration, Salmonova RUs handling decode, and Blackwell GPUs handling pre-fill. Qualcomm's Cristiano Amon announced the Dragonfly data center brand with Snapdragon C details coming at their June investor day. The WSTS raised the 2026 semiconductor TAM forecast by 90% to $1.51 trillion, with Pat noting the market could hit a trillion dollars if memory is excluded entirely. (The Decode) NVIDIA RTX Spark and the Edge AI Push: NVIDIA coordinated with ARM and Microsoft around the RTX Spark at COMPUTEX, with the shared message being that the future of Windows is here. Signal65's Ryan Shrout asked Jensen directly why NVIDIA wants to be in the PC business, given low margins and diminishing returns. Dan frames the answer in the context of devices increasingly becoming mobile data centers, capable of running models at much greater efficiency than cloud delivery. The edge AI conversation is also directly tied to token cost economics: as intelligence delivery moves closer to the device, the cost per token drops significantly. The jury is still out on whether NVIDIA will meaningfully disrupt the PC market, but its influence over OEMs like Lenovo and Dell that depend on it for data center gives it real leverage over SKUs. (The Decode) Token Economics and Frontier Model Cost Pressure: Dan and Pat discuss a substantive shift in how enterprises are thinking about AI consumption costs. Dan argues that "token maxing," the practice of defaulting to the most powerful frontier model for every task, has now effectively peaked, as bills have come due at scale. Companies paying for tokens in volume are starting to question whether they can afford the prices that frontier models actually cost to deliver. Pat pushes back, saying the dynamic is still present, but both analysts agree that the market is moving toward a model where token selection is matched to the job, with Microsoft's MOE approach and thinking models positioned to help CFOs manage that economics story. (The Decode) Continuum Goes Public at Highest Valuation for an AI Platform: Dan notes that Continuum, the Honeywell-spawned quantum company, went public this week at what he calls the highest valuation for an AI platform to date. He flags that IonQ will likely contest that characterization. The broader context is Microsoft entering the quantum conversation with Majorana 2 at Build, a name that has largely been absent from the quantum race, while IBM has received most of the attention. (The Decode) AI CapEx Has Outgrown Cash Flow — Alphabet's $80 Billion Equity Raise: On June 1, Alphabet announced an $80 billion equity capital raise, upsized to $85 billion, structured as $40 billion ATM, $30 billion underwritten, and a $10 billion private placement with Berkshire Hathaway anchoring. Pat frames the questions over CapEx returns as entirely dependent on whether you are an AI boomer or a doomer: if the payback comes, the raise is the right move. If it does not, the math doesn't close. Dan argues the investment is existential, drawing parallels to how infrastructure-first companies have always spent ahead of monetization, and notes that Google's equity is being used as a capital engine that may be more efficient than the debt markets right now. Both analysts flag the downstream implications for Broadcom, MediaTek, and Marvell given the TPU connection. (The Decode) The Network Becomes the AI Platform: Cisco Live 2026: Cisco launched Silicon One P200, the Secure AI Factory with NVIDIA and Spectrum X, AgenticOps, MCP-native automation, Cisco IQ, LiveProtect, and folded Astrix Security and Galileo into Splunk under one control plane. Pat identifies Cisco Cloud Control as the biggest announcement of the entire show, pulling together Catalyst, Meraki, Nexus, Firewall, and WebEx under agentic ops that run natively through MCP, with code running directly on smart switches that have x86 processors. Pat also credits Cisco for establishing Silicon One as a credible chip alternative for hyperscalers capable of taking on Tomahawk and Jericho. Dan frames the long-term opportunity as campus and branch enablement when industrial AI and robotics deployments accelerate, arguing that the numerator of AI's economic impact has barely started, as edge deployment spending has not yet begun. (The Decode) The Flip: Did Microsoft Build 2026 Effectively End the OpenAI Partnership? Pat argues the divorce decree has been filed. MAI-Thinking-1 was built with zero distillation from third-party models offering clean enterprise data lineage, with Maia 200 in production plus Anthropic chip supply, which signals vendor hedging. OpenAI is going all-in on AWS, which means you cannot be married to two people, and the full Build stack covering model, OS containment via MXC, agents via Scout and Agent 365, and context via Microsoft IQ removes every architectural dependency on OpenAI. Dan counters that Microsoft is hedging rather than leaving and predicts the partnership will run through the decade. Enterprise Copilot customers are explicitly showing in data that they demand GPT 5.5, internal benchmarks have not been independently validated, and Microsoft stands to make meaningful money from the OpenAI IPO. (The Flip) Broadcom Q2 FY26 Earnings: Broadcom posted revenue of $22.19 billion, a narrow miss depending on which consensus data set is used, with EPS of $2.44 beating estimates and AI semis at $10.8 billion. Hock Tan declined to raise the $100 billion full-year AI chip target, and the stock dropped 13% in premarket trading. Q3 guide came in at $29.4 billion. Pat calls the miss a timing issue driven by Google's multi-sourcing across Marvell, MediaTek, and Broadcom rather than a fundamental problem. Dan flags that Hock Tan opened the earnings call by accidentally reading from the 2025 print, calling it "not the best moment." Sell-side re-ratings held in the 500s across Jefferies, Mizuho, and Deutsche Bank despite the drop, with Futurum Equities having it at 600. (Bulls and Bears) Hewlett Packard Enterprise Q2 FY26 Earnings: HPE delivered revenue of $10.68 billion, up 40% year over year, and EPS of $0.79, up 100%. Juniper integration and AI servers both outperformed, and all FY26 guides were raised. The stock jumped 19% after hours before settling into a roughly 15% gain, with HPE up 68% over the last month. Pat frames HPE as a value play rather than a volume play, methodically targeting enterprise and sovereign cloud deals where it can maintain profitability, rather than competing for massive NeoCloud volume. Antonio Neri was clear on the call that the profitability pull-forward is a one-shot deal. Pat and Dan will both be at HPE Discover the week after next to interview Neri and the C-suite. (Bulls and Bears) Palo Alto Networks Q3 FY26 Earnings: Palo Alto posted revenue of $3.0 billion, up 31% year over year, beating the $2.94 billion estimate, with non-GAAP EPS of $0.85, beating the $0.79 to $0.81 range. NGS ARR reached $8.1 billion, up 60% year over year, including $1.6 billion from CyberArk and Chronosphere. RPO hit $18.4 billion, up 36%. Both FY26 revenue and EPS guides were raised. Adjusted FCF margin came in at 38.5% TTM, up 430 basis points. The stock jumped 11% immediately after hours, then drifted lower. Pat points to 2,200 platformized customers and 120% net retention as the most important metrics. Dan notes the SaaSpocalypse thesis continues to be wrong. (Bulls and Bears) CrowdStrike Q1 FY27 Earnings and the Proprietary Data Moat Argument: CrowdStrike posted revenue of $1.39 billion with EPS of $1.10 and ARR of $5.51 billion. Net new ARR of $255.8 million se
