Hey, Kai here. This week felt like the moment the AI stack got real across the board. Amazon ditched its “we’re neutral” vibes and rolled out a full vertical—from models to chips. On the hardware front, suppliers showed up with actual, rack-ready alternatives to Nvidia, cooling tubes and all. Meanwhile, Google quietly put an agent in your pocket that will call stores for you and hand back answers, which is going to bend local search in some uncomfortable new ways. And on the software side, the money is following “vibe coding” from gimmick to operations, with backends like Supabase turning AI-written apps into grown-ups. Coffee’s hot—let’s unpack it.
Amazon stops being Switzerland and ships a full-stack AI platform
In a Nutshell
AWS moved from neutral marketplace to opinionated AI platform. At re:Invent, Amazon introduced Nova (its family of first‑party foundation models), Nova Forge (for customizing them), Kiro (a “frontier” coding agent), and Trainium3 chips to power the whole thing. Translation: AWS isn’t just hosting other people’s models anymore—it’s fielding its own model, agent, and silicon layers, tightly wired into developer tooling and governance. This puts Amazon directly against Microsoft/OpenAI and Google, where cloud leadership increasingly means controlling the model, agent workflows, and the hardware underneath. The big near-term question is whether existing AWS customers embrace the convenience and policy controls of a vertically integrated stack—or keep mixing Anthropic, Mistral, Llama, etc., across multiple clouds. Either way, AWS has picked a lane: the cloud as AI co‑pilot, not just compute and storage.
Why Should You Care?
– If you build on AWS, the “easy button” just got easier. Expect faster paths from prototype to production with opinionated defaults for model choice, policy, and security baked in.
– But integration cuts both ways: your architecture will likely become stickier. Plan exit ramps (model abstraction layers, multi-cloud data pipelines) if portability matters.
– Developers get a clearer path for agentic coding. Kiro + Nova + Trainium3 could mean lower latency, better cost curves, and tighter IDE integration for long-running coding workflows.
– Security/compliance teams may prefer first‑party governance over stitching policies across vendors. That could accelerate approvals—and budgets—for AI projects.
– Procurement shifts from “which model?” to “which stack?” Skills in AWS-native AI services will appreciate; generalist model-swapping might depreciate.
-> Read the full in-depth analysis (AWS Nova, Kiro, and Trainium3: Amazon’s Vertical AI Stack Explained)
The non‑Nvidia rack is here (and it’s liquid‑cooled)
In a Nutshell
At SC25, the alternative AI stack showed up in steel, not slideware. MiTAC unveiled a production‑ready, liquid‑cooled AMD Instinct MI355X system in realistic rack layouts, while Gigabyte spotlighted Arm‑based AmpereOne M servers for CPU/control planes. Paired with Arista networks running Broadcom Ethernet fabrics and a maturing ROCm software ecosystem, the message was blunt: you can now order a full, high‑end AI rack without Nvidia or hyperscaler‑only platforms. For enterprises and Tier‑2 clouds that struggled to procure accelerators in 2023–24, this is a phase change—from scrounging cards to deploying coherent, validated systems. It’s not without caveats (software gaps, ops complexity, liquid cooling), but the ecosystem has crossed from “experimental” to “production‑grade alternative.”
Why Should You Care?
– Capacity relief: If your AI roadmap stalled on H100/H200 scarcity, MI355X racks plus ROCm are a viable path to get training and large‑scale inference online.
– Cost/negotiation leverage: A credible non‑Nvidia option improves pricing power with every vendor—from GPUs to networking to managed services.
– Skills investment shifts: Teams will need ROCm proficiency, Arm CPU ops, and liquid‑cooling know‑how. Upskill SREs now or budget for specialist partners.
– Architecture flexibility: Ethernet fabrics (Arista/Broadcom) may simplify supply chains versus proprietary high‑speed interconnects—handy for incremental growth.
– Risk management: Diversifying silicon reduces single‑vendor exposure to supply shocks and roadmap surprises. Start with mixed clusters and workload tiering (fine‑tunes and retrieval-heavy inference first; frontier training later).
– Facilities implications: Liquid cooling is here. Factor in retrofits (CDU placement, heat rejection) and cross‑team drills before the first crate arrives.
-> Read the full in-depth analysis (Non-Nvidia AI Stack at SC25: AMD MI355X and AmpereOne M Go Rack‑Ready)
Google’s agent will call the store so you don’t have to
In a Nutshell
Google is rolling out “agentic calling” in the U.S., turning certain local searches into delegated tasks. Think: tap a button in Search or a Gemini planning flow, and an AI dials nearby businesses, asks structured questions (stock, price, hours), and returns a clean summary. It’s Duplex‑style tech graduating from demo to default plumbing for everyday errands. Under the hood are scripted guardrails, disclosure/consent steps, and integration with local business data. Strategically, this shifts local discovery from browsing and filtering to active information retrieval, with Google’s agent acting as a new front door between consumers and merchants—and a new place for rankings, trust, and analytics to evolve.
Why Should You Care?
– For consumers: This is time back. Instead of phoning five stores, you’ll get a short list with verified answers. Expect it to expand from holiday shopping to appointments, repairs, and returns.
– For local businesses: Your “phone performance” becomes SEO. If you don’t answer, answer poorly, or give inconsistent info, you’ll rank worse in these agent-driven flows. Train staff, tighten scripts, and sync POS/inventory with Google.
– For marketers: Double down on structured data (schema.org), accurate hours/inventory, and call-handling QA. Treat these AI calls like a new acquisition channel with measurable conversion.
– For ops leaders: Expect fewer repetitive calls from humans—but more automated ones. Set policies for consent/disclosure, log interactions, and monitor for miscommunication risks.
– For everyone: Trust will hinge on reliability. Keep receipts (confirmation texts, order holds) and expect appeals/feedback loops when the agent gets it wrong.
-> Read the full in-depth analysis (Google agentic calling: how AI is reshaping local search)
Vibe coding grows up: backends are the real winners
In a Nutshell
“Vibe coding” describes building software by describing what you want—behavior, feel, business outcome—and letting AI assemble the stack. The money says it’s real now. Supabase raised a $100M Series E at a ~$5B valuation by positioning Postgres as an AI‑native backend for these apps. Lovable, an AI dev interface, is widely cited at a $200M revenue run rate. The lesson: the chat UI isn’t the business—the infrastructure that turns AI‑written code into durable, maintainable software is. Supabase has become the database of record for AI‑generated apps from tools like Lovable and Bolt, with auth, storage, and edge functions tuned for AI-centric workflows. The emerging “vibe coding stack” layers model/agent, scaffolding, backend, and ops/observability to turn prompts into production.
Why Should You Care?
– Solo builders and teams can ship full‑stack apps faster than spec‑driven cycles. The constraint moves from “write code” to “define outcomes and guardrails.”
– Budgets shift: Less spend on hand-coding boilerplate, more on managed backends, observability, and runtime governance. Plan for usage‑based costs, not just seats.
– Careers nudge toward architecture over syntax. Prompting, data modeling, and lifecycle design outvalue memorizing framework minutiae.
– Reliability still matters. Wrap AI-generated code with tests, feature flags, and rollbacks. Choose backends (like Supabase) with strong auth, migrations, and edge delivery.
– Avoid vendor lock‑in by keeping the data layer portable (Postgres), using open standards for events/APIs, and tracking model/agent decisions for auditability.
– Executive takeaway: Vibe coding is no longer a toy project. Treat it as a force multiplier—but insist on production discipline from day one.
-> Read the full in-depth analysis (Supabase, Lovable, and the Vibe Coding Infrastructure Stack)
I’ll leave you with a throughline: power is consolidating into opinionated stacks—clouds that own models and chips, hardware vendors that ship full racks, platforms that mediate your calls, and backends that absorb the chaos of AI-generated code. That consolidation can be good—fewer seams, faster shipping—but it raises new questions about lock‑in, governance, and who gets to set the defaults we all live with. Where do you want to be opinionated in your own stack—and where do you want to keep the escape hatches? Hit reply and tell me where you’re doubling down, and where you’re hedging.
