Three new cards, three new operators. Two threads emerged from this scan: AI agents reshaping how products and platforms get discovered, and Apple Silicon's hardware architecture crossing a threshold where trillion-parameter inference leaves the cloud.
Theme 1, AI agents are becoming product-distribution infrastructure
Two findings from different vantage points converge on the same structural shift. AI coding agents are now capable enough to select, configure, and deploy to developer platforms during a live session. That capability makes the agent a word-of-mouth mechanism that operates at the moment of problem-solving.
Val Town CEO Steve Krouse published the May 2026 investor update this week (A coding agent that actively deploys to a platform can become its single largest user-acquisition channel). The headline figure from a new-user acquisition survey: 25% of Val Town's new users in May reported discovering the platform via Claude, making it the single largest acquisition source. Every traditional marketing channel combined accounted for less.
"I was shocked to learn that in the last 30 days, a whopping 25% of new users self-reported learning of Val Town from Claude, dwarfing any other single source of traffic."
The mechanism is specific. When Claude helps a developer write code that needs a deployment target, it recommends and deploys to a platform during the task. Discovery happens exactly when the user has the problem the platform solves. That timing is why the referral rate beats paid channels: no landing page or ad can compete with a recommendation embedded inside the user's active work.
Dave Gerhardt named the buyer-side counterpart of the same shift for B2B marketing (B2B buyers now research vendors in AI tools before any sales contact, making AI-search discoverability a pre-funnel requirement). His "anti-social buyer" framing captures the pattern: B2B buyers now complete most research in ChatGPT, Claude, Reddit, and peer networks before any vendor contact. When an SDR email arrives, the buyer has already formed a competitive landscape from AI-generated answers. Being named in those answers is now a pre-funnel requirement.
"Your buyer already knows about your product. They asked ChatGPT."
The two observations address different markets but follow the same mechanism. AI tools now mediate product discovery before the vendor has any awareness of the prospect. Distribution that worked through search, paid, and outbound continues to work, but it operates after a prior selection stage the vendor cannot see.
- Steve Krouse, A coding agent that actively deploys to a platform can become its single largest user-acquisition channel. AI coding agents that deploy to a platform during live sessions can become the platform's single largest user-acquisition channel.
- Dave Gerhardt, B2B buyers now research vendors in AI tools before any sales contact, making AI-search discoverability a pre-funnel requirement. B2B buyers research vendors in AI tools before any sales contact, making AI-search citation a pre-funnel prerequisite.
Theme 2, Apple Silicon clusters make trillion-parameter local inference practical
The on-device ML story from WWDC 2026 runs deeper than Siri improvements. Angelos Katharopoulos and Awni Hannun shipped JACCL, a collective communication library for MLX, and demonstrated it at WWDC on June 9 (Linking Apple Silicon Macs over Thunderbolt 5 with JACCL achieves 3x LLM inference speed and enables trillion-parameter local inference).
The core result: four M3 Ultra Macs connected via Thunderbolt 5 running MLX with JACCL achieve approximately 3x the inference throughput of a single machine for 27-billion-parameter models. The same cluster runs Kimi K2, a roughly trillion-parameter model, without cloud infrastructure. The same JACCL primitives applied to fine-tuning also showed 3x training throughput.
The mechanism is not primarily software. RDMA over Thunderbolt 5 gives Apple Silicon GPUs near-local-memory latency across machines. JACCL provides the allreduce and allgather operations that tensor-parallel inference requires. Memory capacity scales with the cluster size. A model that no single Mac can hold fits in the combined unified memory of a four-Mac cluster, and the inter-node communication overhead is low enough that wall-clock throughput gains are real.
This matters for practitioners who need large-context inference or fine-tuning without sending data to a cloud provider. The hardware cost is significant (four M3 Ultras), but the model is self-contained, auditable, and runs at a scale that was cloud-only a year ago.
- Angelos Katharopoulos and Awni Hannun, Linking Apple Silicon Macs over Thunderbolt 5 with JACCL achieves 3x LLM inference speed and enables trillion-parameter local inference. Thunderbolt 5 clusters of Apple Silicon Macs with JACCL reach 3x inference throughput and run trillion-parameter models without cloud dependency.