9 March 2026

Open Models, Big Bets, AI at Scale

Big Developments

DeepSeek V4 Arrives: Frontier Performance at Fraction of the Cost

DeepSeek V4 launched as a ~1 trillion parameter multimodal model supporting text, image, and video generation with a 1M token context window. Priced at roughly $0.14 per million input tokens, about 1/20th the cost of GPT-5, it was optimised for Huawei Ascend chips.

Why it matters
If your team is running AI workloads and hasn't done a cost audit lately, this is your nudge. V4's pricing makes the "just use GPT for everything" approach look like ordering champagne when the house wine is perfectly drinkable. For adviser-facing tools at Benchmark, cheaper models mean more room to experiment without blowing the budget.

Meta Signs $100B AMD Chip Deal, Building Multi-Vendor AI Infrastructure

Meta announced a deal potentially worth up to $100 billion to purchase AMD MI450 GPUs covering six gigawatts of data centre capacity. The deal follows Meta's separate Nvidia commitment days earlier, signalling a deliberate multi-vendor chip strategy.

Why it matters
Think of it like not putting all your ISA in one fund. Meta's diversifying its AI supply chain, and the same logic applies to product teams picking model providers. If you're building features on a single AI vendor, you're one pricing change away from a very awkward sprint planning session.

Apple and Google Announce Multi-Year Deal to Rebuild Siri on Gemini

Apple and Google entered a multi-year partnership under which the next generation of Apple Foundation Models will be built on Google's Gemini models and cloud technology. The new Siri targets a March 2026 launch with iOS 26.4, featuring on-screen awareness and deep cross-app context.

Why it matters
Apple basically said "we'd rather ship something good now than wait to build it ourselves." That's the build vs. buy conversation every product team has on repeat. For wealth management tech, there's a lesson here: sometimes licensing a solid AI capability beats spending 18 months building a mediocre one in-house.

Product Thinking

Amy Mitchell: Your Product Is Now Evaluated by AI Before Humans See It

Product manager Amy Mitchell argues that products are now evaluated twice: first by AI systems during search and discovery, then by human buyers. When product information isn't structured and machine-readable, AI fills the gaps with outdated framing or competitor comparisons. Treating 'product explainability' as a PM discipline is the fix.

The Deployment Reality Check: Agentic AI Is Struggling to Prove ROI

March opens with a clear industry shift: benchmark races have given way to harder questions about whether agentic AI actually performs reliably in production. PwC notes that many 2026 agentic deployments haven't delivered measurable value, and organisations now need proof points tied to real business outcomes, not demo impressions.

Tools to Explore

Bagel AI: Evidence-Based Roadmap Prioritisation Tied to Revenue

Bagel AI extracts signals from sales calls, support tickets, CRM updates, and customer notes, then connects those signals directly to revenue impact and churn risk. Unlike roadmapping tools that rank by votes, Bagel ranks by business consequence. Worth a look for any product team struggling to tie qualitative feedback to ROI in stakeholder conversations.

Apollo.io AI Assistant: Agentic GTM Execution in a Single Platform

Apollo.io launched its AI Assistant out of beta with nearly 20,000 weekly active users. The assistant executes complex workflows like prospect discovery, research, outreach, and deal optimisation using natural language. Product teams building agentic features into sales or CRM workflows should examine this as a live benchmark for embedded agent UX.

What's up with Schroders?

Schroders shares stall near 585p as March scheme document and April shareholder vote dates confirmed

Schroders shares have settled around 585p, well below Nuveen's 612p-per-share offer, reflecting the long road still ahead. New regulatory filings confirm the scheme document and proxy form are due in March, with the court meeting and general meeting set for April 2026. The deal remains on track to close in Q4 2026, pending shareholder and regulatory approval.

Feelgood Features

Gates Foundation and OpenAI Launch $50M Horizon1000 to Bring AI to African Clinics

OpenAI and the Gates Foundation committed $50 million to Horizon1000, a pilot starting in Rwanda that aims to deploy AI tools across 1,000 primary healthcare clinics in sub-Saharan Africa by 2028. The region faces a shortfall of nearly 6 million health workers. Tools will support clinical navigation, admin burden reduction, and patient triage.

AI System DeepRare Brings New Hope for Patients With Rare Diseases

Researchers published findings in Nature on DeepRare, an AI system that generates ranked diagnostic hypotheses for rare diseases using clinical data, genetic information, and medical literature. With over 7,000 recognised rare diseases, even experienced clinicians struggle to connect unfamiliar symptom patterns. DeepRare provides reasoning traces alongside each hypothesis, linking conclusions to verifiable evidence.

If you have any feedback or issues with this newsletter, please do speak to your nearest Bench and they will be happy to assist.