Friday, July 3, 2026

The through-line isn't a model launch — it's how the AI buildout is reshaping org charts and headcount. Microsoft stands up a $2.5B "Frontier Company" of embedded engineers (days after Amazon's own move), neocloud Together AI raises $800M as open-model inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ demand surges, and the June jobs data lands with AI now the single most-cited reason for U.S. layoffs four months running. On the engineering side, the actionable item is quieter: Akamai's MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ database flaws are a reminder to audit any MCP server wrapping a database.

TLDR

Microsoft launches the $2.5B "Frontier Company" — ~6,000 forward-deployed engineers embedded inside customers, days after Amazon's $1B FDE unit (Jul 2) Business
Together AI raises $800M Series C at $8.3B valuation (Aramco Ventures-led) as open-model inference demand surges (Jul 1) Infra
June jobs: AI is the #1 cited layoff reason for a 4th straight month (14,029 cuts in June, 101,743 YTD); BLS payrolls rose just +57k (Jul 2–3) Hiring
TwelveLabs raises $100M Series B for "video superintelligence," signs AWS Trainium deal; ships Marengo 3.0 embedding model (Jul 1) Customer AI

Action Items

Things your engineering team should look at or do this week.

Audit any MCP server that wraps a database. Akamai disclosed three back-end MCP flaws — SQL injection in Apache Doris MCP (CVE-2025-66335, <0.6.1), an unauthenticated metadata-exfiltration bug in Alibaba's RDS MCP (vendor declined to fix — still live), and a potential takeover in Apache Pinot's MCP. Check input sanitization and auth on your own MCP endpoints. (Akamai)
Re-price open-model inference against neoclouds. Together AI's raise comes with ~500MW of committed capacity and disclosed >$1.15B annual bookings; if you're defaulting to closed OpenAI/Anthropic APIs for open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ workloads (DeepSeek, MiniMax, Kimi), it's worth re-benchmarking cost per tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ on a neocloud. (TechCrunch)

Models & Tools

TwelveLabs raises $100M Series B; Marengo 3.0 + Pegasus 1.5 video models (Jul 1) Customer AI Data — Co-led by NEA and NAVER Ventures (with Amazon, Radical, Index, Red Bull Ventures), the round funds a shift from video understanding to a full-stack "agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ intelligence system for video." Marengo 3.0 is positioned as a top video-embedding model (search across sound, speech, and motion over time); Pegasus 1.5 turns footage into structured data (scene boundaries, entities, temporal segments). It also signed an AWS Trainium deal — relevant if you're building video RAGRetrieval-augmented generationFetching relevant documents at query time and giving them to the model, so its answer draws on them.Lewis et al. 2020 ↗/search over large archives. (GlobeNewswire | AI Weekly)

Security

Akamai: three MCP back-end database flaws, one left unpatched (disclosed June, presented at x33fcon) Security Agents Data — Akamai's research ("One Is a Fluke, 3 Is a Pattern") documents a class problem in database-connected MCP servers: unsanitized SQL, missing auth, and unauthenticated data exposure. The Apache Doris SQLi has a CVECommon Vulnerabilities and ExposuresA public catalog that gives each known security vulnerability a unique ID, a short description and at least one reference.NIST ↗ and a fix (0.6.1+); the Alibaba RDS MCP metadata-exfiltration bug was marked "not applicable" by the vendor and remains in the codebase. Broader scans in the same window (VIPER-MCP → dozens of CVEs; ~40% of remote MCP servers reportedly unauthenticated) make this a good week to lock down MCP endpoints. (Akamai | The Register)

Industry & Business

Microsoft launches the "Frontier Company" — a $2.5B forward-deployed engineering unit (Jul 2) Business — Microsoft is embedding ~6,000 engineers (mostly existing staff) inside customers to build, deploy, and run AI systems on-site, led by Rodrigo Kede Lima (ex-president, Microsoft Asia); initial clients include Unilever and Novo Nordisk. It follows Amazon's $1B FDE commitment two days earlier and mirrors OpenAI's TPG-backed Deployment Company and Anthropic's own unit (both May). The whole hyperscaler cohort is now betting that deployment, not just model access, is the enterprise moat. (TechCrunch | GeekWire)
Together AI raises $800M Series C at $8.3B valuation (Jul 1) Infra — The neocloud (Nvidia GPUGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗ clusters for training/deploying open models) roughly 2.5x'd its valuation from 16 months ago. Round led by Aramco Ventures with Vista, General Catalyst, Nvidia, and others; >$1.15B annual bookings, customers including Cursor and Cognition, and ~500MW of committed compute earmarked for ~50x capacity growth over five years. A marker of how much demand is pooling into open-model inference. (TechCrunch | The Next Web)
June jobs: AI leads layoff reasons for a 4th straight month (Jul 2–3) Hiring — Challenger, Gray & Christmas reports 45,849 June job cuts (down 53% from May), but AI was the single most-cited reason for the fourth consecutive month — 14,029 cuts (31%) in June, 101,743 YTD (~23% of all 2026 cuts). Tech drove ~a third of H1 cuts (139,156 YTD, +83% YoY). Separately, BLS reported June nonfarm payrolls up just +57k with unemployment at 4.2% — a soft print for planning conversations. (Challenger | BLS)

Policy & Safety

UN's Independent International Scientific Panel on AI issues its first preliminary report (Jul 1) Policy — The newly-formed UN panel's inaugural report warns that AI capability is accelerating faster than the governance and safety rules meant to contain it — the first output of the body created to give member states a shared scientific baseline (an "IPCC-for-AI" model). Non-binding, but a signal of where international coordination is heading ahead of the EU AI Act's Aug 2 GPAI deadline. (UN News)

A capital-and-labor day.