What 27 Sessions at the World’s Biggest AI Conference Revealed About Where the Money Flows Next

Disclaimer: I hold small positions in some securities mentioned in this article. This is not financial advice. Everything here is for research and educational purposes only. Do your own due diligence before making any investment decisions.

I spent two days at HumanX 2026 in San Francisco covering 27 sessions as a media representative. My background is in supply chain, ERP operations, and fintech — and I build automated trading systems and web applications for financial markets. I run Trade with Harp, a paid investment research and trading community.

What follows is what I took away from the conference specifically for investors. Not a session recap. An investment map.

The framework: the 5-layer stack

The opening session at HumanX did something useful that most AI coverage doesn’t — it mapped the entire AI investment universe into five discrete layers. Every company and thesis in this piece maps to one or more of these layers.

Layer 1 — Chips: GPUs, CPUs, memory silicon. Structural, multi-year demand. Layer 2 — Infrastructure: Data centers, cloud compute, networking. Capex-intensive, contractually backed. Layer 3 — Models: Foundation model providers. Competitive, high burn, uncertain moats. Layer 4 — Applications: Enterprise software, SaaS, agents. Mixed — watch for disruption vs. survival. Layer 5 — Energy: Power, nuclear, grid. Underappreciated, long-duration thesis.

The core thesis confirmed across two days of sessions: Layers 1, 2, and 5 have the most durable investment case. Layer 3 is competitively unstable. Layer 4 is bifurcating — some companies will compound, others will get disrupted. The conference consensus was that almost everyone is investing in layer 3 and 4 while the real structural opportunity sits at layers 1, 2, and 5.

Tier 1 — Infrastructure (highest conviction from the conference)

CoreWeave (CRWV)

What HumanX said: Peter Salanki, CoreWeave’s CTO, explained their moat plainly — they built AI-first architecture from scratch rather than retrofitting hyperscaler stacks. AI workloads couple many servers into single tasks, making partial failures far more impactful than in traditional cloud. That’s why purpose-built infrastructure has a structural advantage that legacy cloud providers struggle to replicate quickly.

What the market says right now: Revenue grew 168% to $5.1B in 2025. Management is guiding 235%+ growth in 2026 — implying roughly $12B revenue. Contracted revenue backlog sits at $66.8B. This week alone: a $21B deal with Meta (April 9) and a $6.8B deal with Anthropic (April 10). An $8.5B debt financing facility closed March 31, secured against GPU clusters and the Meta backlog. The stock is down roughly 39% from its 52-week high despite that backlog growth — macro headwinds and construction delay concerns are weighing on sentiment. They plan to spend $30–35B on infrastructure in 2026 alone. The main risk: 67% of 2025 revenue came from Microsoft — customer concentration that the Meta and Anthropic deals are actively diversifying.

The tension: enormous backlog versus enormous capex. Profitable if demand holds. Vulnerable if AI efficiency improvements reduce compute demand faster than expected.

Watch for: customer concentration declining below 50% for any single client, margin improvement as infrastructure matures, and Rubin GPU deployment timelines.

Micron Technology (MU)

What HumanX said: Sumit Sadana, Micron’s CBO, described a step-change in data center memory demand that began in late 2025, constrained by multi-year supply lead times. Micron increased planned CapEx to over $25B and is actively seeking multi-year customer co-commitments to share demand risk. Power per bandwidth is the primary bottleneck driving LPDDR adoption, HBM improvements, chiplet partitioning, and 3D stacking.

What the market says right now: Q2 FY2026 results (March 18) were described as blowout — management’s guidance annihilated Wall Street expectations. The stock has retreated roughly 30% since that report on concerns about CapEx cycle sustainability and margins. Memory supply deficit is estimated to last until 2030. HBM — high bandwidth memory — is the specific product category driving AI demand, and Micron is one of the few companies that can supply it at scale.

The tension: strong fundamental demand story versus market concern about the CapEx cycle.

Watch for: HBM revenue as a percentage of total, CapEx guidance trajectory, and any signal of demand pull-forward or slowdown from hyperscalers.

NVIDIA (NVDA)

What HumanX said: Bryan Catanzaro from NVIDIA described four scaling drivers simultaneously increasing compute demand — pre-training, post-training, deployment compute cycles, and agents. All four are additive. NVIDIA’s co-design approach spans from transistors to frameworks. The Nemotron project is used internally to learn about AI system requirements and guide future hardware design.

What the market says right now: The Vera Rubin platform (R100/R200) is now in deployment — fabricated on TSMC’s N3P process, featuring HBM4 delivering 22 TB/s of memory bandwidth. The CoreWeave/Meta $21B deal is partly a Rubin GPU access play. NVIDIA remains the tax collector of the AI era, benefiting from every GPU CoreWeave deploys. Arista Networks (ANET) was described in multiple sessions as the mandatory networking plumbing for 100,000-GPU clusters.

The tension: NVIDIA’s valuation already reflects much of the opportunity. Upside depends on whether AI compute demand continues to scale as expected — which every session at HumanX suggested it will.

AMD (AMD)

What HumanX said: Mark Papermaster, AMD’s CTO, highlighted diverse workload demand including CPUs, GPUs, inference, and agentic flows. AMD and Micron are co-designing open standards for compute-memory integration, which signals AMD is a serious player in the AI silicon layer.

The thesis: AMD is the NVIDIA alternative play — benefits if hyperscalers diversify GPU sourcing to reduce dependency on a single supplier. Lower valuation than NVIDIA, higher execution risk.

Arm Holdings (ARM)

What HumanX said: Will Abbey, Arm’s EVP, described Arm’s deliberate move from IP licensing to compute subsystems and full chips, driven by continuously running agentic workloads that demand power efficiency above all else. 350 billion Arm cores have been shipped. Performance-per-watt is the metric that matters now.

The thesis: as agents run 24/7, token economics explode. Power efficiency becomes a competitive moat at the infrastructure layer. Arm is uniquely positioned for inference at the edge. Less discussed than NVIDIA but structurally significant.

Tier 2 — Enterprise software (watch the bifurcation carefully)

The session titled “SaaS Pocalypse: Is AI About to Eat Enterprise Software Whole?” didn’t provide a clean answer. But several sessions around enterprise AI adoption gave clear signals about which companies are positioned to survive the transition — and which are exposed.

ServiceNow (NOW)

What HumanX said: Pat Casey, ServiceNow’s CTO, described AI as a multiplier of engineering output. ServiceNow is building internal AI tools on their own platform to remove bottlenecks — eating their own cooking. Risk flagged: uneven productivity gains across engineers and a potential junior talent pipeline issue as AI reduces entry-level training opportunities.

Current market signal: Q4 revenue $3.6B, 21% year-over-year growth. 2026 subscription revenue forecast of at least $15.5B.

The thesis: ServiceNow is the workflow orchestration layer for enterprise AI. Every HumanX session on agentic AI confirmed that orchestrating agents across systems is the hardest unsolved problem. ServiceNow’s integration platform becomes more valuable in that world, not less.

Salesforce (CRM)

Current market signal: Revenue $10.3B, 9% year-over-year growth. Agent Force AI product grew annual recurring revenue 330%.

The thesis: lower forward P/E than peers, CRM market share intact. Agent Force is a real product with real traction. 330% ARR growth is not marketing — it’s revenue.

Databricks (private — watch for IPO)

What HumanX said: CEO Ali Ghodsi made the argument that will define enterprise AI investment for the next decade: current models are sufficiently capable, but they fail in enterprises because they lack context. The bottleneck isn’t the AI. It’s the data infrastructure and governance around it. He expects enterprise adoption to take five to ten years.

The thesis: Databricks is the data layer bet. Every enterprise AI deployment eventually requires clean, governed, accessible data. Databricks’ Lakehouse model is positioned as the default infrastructure for that requirement. Pre-IPO — worth watching for listing.

Tier 3 — Fintech and financial infrastructure

Two specific sessions gave high-signal investment intelligence for fintech.

Visa (V)

What HumanX said: Rajat Taneja from Visa described the company adapting its fraud, privacy, and safety stack for agentic commerce — new identity systems, permissioning frameworks, and tokenized credentials that allow AI agents to transact on behalf of consumers safely. The shift will be gradual, but Visa’s infrastructure is the trust layer that makes it possible.

The thesis: as AI agents begin making purchases autonomously, payment authentication and fraud detection become more critical, not less. Visa’s moat deepens in an agentic commerce world rather than being disrupted by it.

Fintech security plays (PANW and others)

What HumanX said: Jonathan Levin of Chainalysis described AI dramatically lowering the barrier for financial fraud — enabling impersonation, automation, and scale that wasn’t previously accessible to lower-skill actors. Defense requires proactive threat hunting, intelligence-sharing networks, and AI agents for evidence collection. The financial crime and compliance sessions both confirmed that the regulatory and litigation pressure forcing governance investment in financial services is real and accelerating.

The thesis: AI-enabled financial fraud is structurally accelerating. Companies providing financial crime detection, AML tools, and identity verification infrastructure are benefiting from a forced adoption cycle. Public plays include Palo Alto Networks (PANW) and the broader security infrastructure space.

Tier 4 — Energy (long-duration, under-owned)

What HumanX said: Al Gore’s session addressed AI’s energy consumption directly. The infrastructure session identified energy as the dominant, hard-to-scale constraint on AI growth — ahead of chips and memory. CoreWeave’s entire buildout is gated by power availability. The silicon shift session confirmed that data center power requirements are on a trajectory that renewable energy sources alone cannot meet reliably.

The thesis from the room: data centers being planned and funded right now need reliable baseload power. Nuclear is the conversation nobody wants to have publicly but everyone is having privately. The sessions confirmed this without anyone saying it directly.

Relevant positions: uranium producers and nuclear operators including Cameco (CCJ) and NuScale (SMR); power infrastructure including Vistra (VST) and Constellation Energy (CEG); and Vertiv (VRT) for liquid cooling in high-density AI data centers — mentioned specifically as a CoreWeave infrastructure beneficiary.

Key risks heard consistently across 27 sessions

The reliability problem is unsolved. Every agentic AI session flagged reliability as the primary constraint on enterprise adoption. If reliability doesn’t improve, the five-to-ten year adoption timeline extends further and application-layer valuations get challenged.

Customer concentration. CoreWeave derives 67% of revenue from Microsoft. Watch for diversification — the Meta and Anthropic deals are the right direction but concentration remains a risk.

Model efficiency risk. If next-generation models require significantly less compute per task — smaller models doing more — infrastructure demand assumptions get challenged. This was mentioned as the primary long-term structural risk to the infrastructure thesis.

Governance and compliance lag. Multiple sessions confirmed most enterprises lack basic AI governance. This slows adoption timelines and creates liability risk for early-moving enterprise software companies that deploy before governance frameworks are ready.

The supply chain gap. The $15–20T supply chain industry has almost no meaningful AI penetration. This is either a massive untapped opportunity or evidence of structural barriers that will slow the broader enterprise AI adoption thesis from the demand side. I wrote about this gap separately — it’s the most underreported story from the conference.

What the conference confirmed versus what it challenged

Confirmed: infrastructure (chips, data centers, energy) is where structural demand sits. Enterprise adoption is slower and harder than the narrative suggests — five to ten years is a realistic timeline, not pessimism. Agentic AI is real but reliability is the gating factor. Private enterprise data is the durable moat, not foundation model access. Compliance and governance infrastructure will compound for companies that build it early.

Challenged: the application layer is more competitive and more vulnerable than headline coverage suggests. The label “AI-powered” means almost nothing without evidence of production reliability. Small and mid-size businesses are almost entirely excluded from current AI tooling — which limits the addressable market assumptions underlying many application-layer valuations.

The conference was bullish on the direction of AI. It was much more measured — in the rooms where the actual builders and operators were talking — about the pace.

That gap between direction and pace is where most of the current mispricing lives.

Jinlu Wang builds automated trading systems and web applications for financial markets. She has a background in supply chain, ERP implementation, and enterprise operations, and covered HumanX 2026 in San Francisco as a media representative. She runs Trade with Harp, a paid investment research and trading community.