AI infrastructure fragments — how this changes your tech stack

Summary

Nvidia split reporting — hyperscale clouds and enterprise direct purchases now separate categories, signaling pricing and supply model divergence
OpenRouter doubled valuation to $1.3B — aggregators routing across 400+ models grow faster than model labs themselves
Twenty CRM (44k stars, $5.5M funding) — open-source Salesforce alternative proves vendor independence is moving to the application layer

Core takeaway: AI infrastructure is no longer vertically integrated. It’s fragmenting into layers — hardware, models, applications — and each layer wants independence.

1. Nvidia changed financial reporting — hardware layer shows fragmentation

What changed. Nvidia Q1 2026 results ($81.6B revenue, +92% YoY in data centers) introduced new reporting structure: data center sales now split into two categories — hyperscale and AI clouds (AWS, Google Cloud, Azure) and industrial/enterprise (ACIE — direct purchases). This isn’t accounting cosmetics — it’s Nvidia’s acknowledgment that hardware purchasing channels and pricing models are now fundamentally different for two customer groups.

Why it matters. If you buy AI compute through a cloud provider (AWS, Azure, Google Cloud), your pricing curve, SLAs, and availability differ from those of an enterprise buying hardware directly from Nvidia. Nvidia now reports these two segments separately — meaning they see: cloud customers grow at one rate, direct enterprise purchases at another. If your strategy is “everything through AWS”, you’re competing with enterprises that buy direct and pay differently.

What to do this month.

Audit your AI infrastructure cost breakdown — what you pay through cloud (AWS/Azure/GCP) vs. what direct deployment would cost
If you spend $500k+/year on AI compute via cloud, request competitive quotes from alternative providers (CoreWeave, Lambda Labs) — Nvidia’s ACIE segment exists because this alternative is becoming real
Identify whether your workloads are “hyperscale” (elastic, burst-heavy, shared GPU acceptable) or “enterprise” (predictable, dedicated, latency-sensitive) — first benefits from cloud, second potentially doesn’t

What I expect.

Nvidia Q2 2026 earnings (August) will reveal which segment grows faster
AWS/Azure/GCP pricing pressure — if enterprise direct grows faster, hyperscalers will either cut margins or offer “enterprise tier” with dedicated GPU guarantees
6-12 months will produce public case studies of companies migrating from cloud to dedicated AI hardware
Nvidia’s forward guidance will split by category, confirming the divergence is structural

2. OpenRouter $1.3B — aggregators beat vendor lock-in

What changed. OpenRouter — AI model aggregator founded in 2023 — raised $113M Series B at $1.3B post-money valuation (led by CapitalG, Alphabet’s growth fund). That’s 2.4x from $547M a year ago. OpenRouter provides routing across 400+ models (Anthropic, OpenAI, Google, xAI, DeepSeek, open-source), processes 100 trillion tokens monthly (5x growth in 6 months), and serves 8 million users. Sacra estimates OpenRouter reached $50M ARR by early 2026, up from $10M in October 2025.

Why it matters. If an aggregator that doesn’t build models, just routes between them, doubles valuation yearly and processes 100 trillion tokens monthly — it means enterprises systematically choose “best model for task” architecture over “one vendor for all”. OpenRouter customers pay for the right NOT to be locked into OpenAI or Anthropic. If you’re paying OpenAI $50k/month for GPT-4 today but 30% of your workload could run on DeepSeek (5x cheaper), the OpenRouter business case is obvious.

What to do this month.

Build an API audit: which tasks in your AI stack use which model, and whether alternatives exist with better cost/performance
Test OpenRouter, Fireworks, or Baseten with 10% of your workload in parallel with existing API — compare latency, accuracy, cost
Request a contract review from your current provider (OpenAI, Anthropic, Google) — if your contract blocks dual-source, you’re locked in exactly when the market is unlocking
If you built in-house model routing logic, consider replacing it with OpenRouter — their 400+ model library grows faster than you can maintain

What I expect.

Fireworks and Baseten will announce similar valuations in 2026 H2 — if all 3 aggregators are unicorns, that confirms routing as a permanent stack component
OpenAI or Anthropic will attempt to acquire an aggregator or build “official routing API” — but that’s strategic contradiction (they want lock-in), so probably won’t happen
12-18 months will see “tier 2” aggregators — niche routing platforms for specific industries (healthcare, finance) with compliance-first routing logic
Cost per million tokens will continue falling, but routing margin will rise — aggregators win not on price but on orchestration value

3. Twenty CRM 44,000 stars — open source moves to the application layer

What changed. Twenty — Paris-based open-source CRM founded in 2023 — reached 44,000 GitHub stars and $5.5M funding across two rounds. Led by Runa Capital, investors include HubSpot founder Dharmesh Shah, former Pipedrive CEO Sergei Anikin, and Front founder Mathilde Collin — CRM industry veterans backing the open-source alternative to what they themselves built. Twenty offers self-hosting free (AGPL-3.0 license) or cloud from $9/user/month. Tech stack — React frontend, NestJS backend, PostgreSQL — plus native MCP server for AI agent integration (the new standard Anthropic introduced).

Why it matters. Salesforce Professional licenses cost thousands of euros per year per user. Twenty offers similar functionality at one-tenth the price or free. More importantly — Twenty is the first serious open-source CRM with built-in AI agent integration. If you plan AI agents that read and write CRM data directly (automated lead processing, task execution, contact enrichment), Twenty offers this out of the box. HubSpot and Salesforce are only now adding AI features on top. Data ownership — self-hosted CRM means customer data stays in your infrastructure, solving GDPR requirements without additional contracts.

What to do this month.

Pull your current CRM annual cost — if you pay >€10k/year for Salesforce or HubSpot Enterprise, Twenty’s business case is real today
If your CRM contract ends in the next 6 months, request a competitive review — Twenty’s existence means you now have negotiating leverage you didn’t have before
Set up a Twenty self-hosted test instance with 10 test contacts and 1 AI agent integration — verify functionality meets your needs
If you use AI agents already (Anthropic Claude, OpenAI), ask your IT lead: does our CRM support MCP server standard? If no — you’re one or two years behind

What I expect.

2026 H2 will see 2-3 more serious open-source CRMs (potentially from Strapi or Supabase teams) — the market learned that open source + cloud service is the winning formula
Salesforce and HubSpot will announce their MCP server integrations by end of 2026 — but at premium prices, contradicting Twenty’s approach
Twenty will raise Series C in 12-18 months at $100-200M valuation — open-source commercialization grows faster than traditional SaaS
For mid-market companies (50-500 employees), Twenty will become the “safe second choice” — either you stay with current CRM or move to Twenty; there’s no third option

Today’s picture

AI infrastructure fragments across three layers simultaneously, and it’s not a technical coincidence — it’s a structural shift changing how enterprises buy, integrate, and present technology stacks. Nvidia’s reporting split signals the hardware layer no longer has one pricing model — hyperscale clouds and enterprise direct are two different economies. OpenRouter’s $1.3B valuation proves aggregators grow faster than model builders in the inference layer, because enterprises choose vendor independence. And Twenty CRM’s 44,000 GitHub stars with HubSpot and Pipedrive founders backing it confirms the same shift in the application layer — even industry veterans support open source over what they themselves built.

Summary table:

Layer	Shift	Signal
Hardware	Nvidia split hyperscale/enterprise reporting	Pricing and supply models diverge
Models	OpenRouter $1.3B (2x in a year)	Aggregators grow faster than labs
Applications	Twenty CRM 44,000 stars, $5.5M funding	Open source challenges Salesforce and HubSpot

Three questions for leaders:

Is your AI stack today “one vendor, one API” — and if so, do you know what migration to aggregator-based routing with 2-3 alternative models would cost?
Did your team evaluate open-source CRM alternatives at the last contract renewal — or just extend Salesforce/HubSpot because “it was convenient”?
Does your IT director know what an MCP server is and why Twenty CRM has it built in — and whether your current providers support it?

Summary

1. Nvidia changed financial reporting — hardware layer shows fragmentation

2. OpenRouter $1.3B — aggregators beat vendor lock-in

3. Twenty CRM 44,000 stars — open source moves to the application layer

Today’s picture

Sources