HubSpot rebuilds its CRM around context — and plugs into ChatGPT Ads
Three vendors moved in the same direction today: the product is no longer the prompt box, it is the context behind it.
HubSpot used its Unbound conference in Boston to ship what it calls its most foundational release in years — a Smart CRM rebuilt around a continuously updated context layer. The company says the platform now syncs calls, emails and meetings into a single view, scores how complete that context is, and flags the gaps that quietly degrade AI output. Breeze Assistant stops being a question-answering box and becomes a work router: describe an objective and it selects specialized agents to draft proposals, reports and campaign plans from CRM data. A new Marketing Studio points at opportunities most teams only notice late — weak visibility in answer engines, underperforming audience segments, leads nobody followed up on.
For HubSpot, the pitch is that context is the moat. Conventional AI tools fail in sales and marketing because they cannot see what the business actually knows, and whoever holds the customer record holds the grounding layer that agents — including advertising agents — have to rent.
That is why OpenAI's same-day advertising announcements matter more than they read. Sponsored Agents, now in testing with select US advertisers, let a user start a clearly labeled conversation with a business's agent after clicking an ad in ChatGPT. ChatGPT Ads also gained its first CRM partner in HubSpot and its first ecommerce partner in Shopify, with Sponsored Agent conversations kept separate from ChatGPT's own answers and from the user's original thread.
HubSpot's supporting numbers are its own: early Marketing Studio users created 81% more campaigns, Breeze Assistant customers generated 2.2 times more leads, and the company cites nearly 3.6 times more marketing-qualified leads platform-wide. No methodology was published, so treat them as vendor claims until an independent audit exists. The strategic read holds anyway — the CRM is becoming the context store that ad agents plug into, and HubSpot just got there before Salesforce.
Local runners got a much cheaper way to hold Alibaba's Qwen3.8-Flash-Next in memory. A new GSQ-RCO quantization on Hugging Face cuts the model from roughly 80–95 GB down to 68–76 GB while staying near baseline quality, and its speed-focused Q2_0 variant claims 3.4 times the prompt throughput and 1.9 times lower end-to-end latency than IQ2_XS at a slightly smaller file size. The trade is small: 89.07 against 89.16 on the release's task average.
The community reaction has been the useful part. Runners on dual 3090s report the build still lacks tensor-parallel and MTP support, so it can lose to a W4A16 alternative on multi-GPU rigs, and a separate thread shows most of Qwen3.8-Flash-Next's KV cache can be parked in system RAM with little decode slowdown. Together that is a 24 GB-class story, not a data-center one — the open-weights gap closes at the quantization layer long before it closes at the training layer.
Vishal Sikka's Hang Ten Systems raised another $53 million, four months after launch. The round was led by Temasek-backed Xora and takes total funding to $85 million, with Aramco Ventures, Intel's Lip-Bu Tan, Micron's Sanjay Mehrotra and Yahoo co-founder Jerry Yang participating. The pitch is enterprise software built and modernized with AI: Sikka says teams of two to four people now do work that previously took about 30, with customers including Saudi Aramco, Siemens Energy and Fresenius Kabi.
That is the bull case for AI eating IT services, and it is a self-reported one — Hang Ten promises a tenfold improvement in cost, speed or both, while final quality checks and certification still sit with customers or independent third parties.
What to watch: whether HubSpot publishes methodology for those multiplier claims, and whether the CRM integrations push OpenAI's ads product past the "sponsored" trust problem.
Which breaks first — the CRM vendors' context moats, or the services firms' headcount model? Tell us in the comments.
