The most interesting AI lab on Monday was a 173-year-old information company. Thomson Reuters says it spent $40 million—not several billion—to turn an open foundation into a model specialized for legal, tax, and professional work. That leads AI news today because it points to a different competitive map: valuable data, expert feedback, and verification may let established companies build models that outperform their size in narrow, expensive workflows. Elsewhere, the money was considerably less restrained. Alibaba priced a $10.2 billion share placement entirely for AI, while Gartner lifted its 2026 semiconductor-revenue forecast to $1.6 trillion. OpenAI’s GPT-5.6 arrived in AWS’s Kiro coding agent, Google Cloud deepened its Verizon partnership, and NVIDIA was reportedly discussing an investment in Perplexity at a valuation above $30 billion. If you run an AI roadmap, today’s useful question is not simply “Which model?” It is “Which advantage can we actually own?”

AI news today: Thomson Reuters builds a $40 million specialist model

Thomson Reuters launched Thomson, its first proprietary large language model, after investing $40 million in talent and compute. The company says it began with an open-source foundation, then applied proprietary content, mid-training, post-training, and feedback from subject-matter experts. Its August 24 announcement says less than 10% of its content has been used so far and describes early performance as comparable with frontier models across selected professional tasks.

Those comparisons remain company claims pending broader external evaluation. Still, the strategic point is real: Thomson Reuters now controls a model, not only the data and applications around one. A small version will be offered as open weights for academic, non-commercial use, while the first product deployment is planned for Tabular Analysis in CoCounsel Legal. Watch the technical report, independent testing, licensing terms, and evidence that specialization improves accuracy rather than merely matching an internal benchmark.

Alibaba turns a record-sized share placement into an AI budget

Alibaba priced 710 million new Hong Kong shares at HK$112.70 each, raising HK$80 billion—about $10.2 billion. The company’s filing says 100% of the net proceeds will fund full-stack AI capabilities, including infrastructure. The placement, offered to non-U.S. investors, is expected to close August 26 subject to customary conditions.

Monday’s market reaction reflected the trade-off: shareholders get a larger AI war chest and immediate dilution. The scale also turns a model strategy into a financing strategy. Alibaba must convert capital into chips, cloud demand, and durable adoption of Qwen and related products. That pressure is not unique to China; our August 21 brief followed similarly aggressive financing around U.S. AI infrastructure. The next proof point is return on that spending, not another large model announcement.

Gartner now forecasts a $1.6 trillion semiconductor year

Gartner raised its forecast for worldwide 2026 semiconductor revenue to $1.555 trillion, up 92% from $809 billion in 2025. Its August 24 projection assigns $837.3 billion to memory alone and expects memory to account for 54% of industry revenue this year. Gartner also estimates that AI data centers will represent 36.5% of semiconductor revenue in 2026 and more than 53% by 2030.

This is a forecast, not booked revenue, and the astonishing growth rate partly reflects memory-price inflation. For buyers, that distinction does not make the bill imaginary. High-bandwidth memory, DRAM, and storage are becoming strategic constraints alongside accelerators and power. Google’s expanding custom-chip ecosystem, covered in our August 20 edition, is one response. The practical move is to model memory and networking exposure separately instead of treating “GPU cost” as the whole infrastructure budget.

GPT-5.6 joins Kiro’s spec-driven coding workflow

OpenAI made the GPT-5.6 Sol, Terra, and Luna models available in Kiro, AWS’s software-development agent. Kiro organizes work into requirements, technical designs, executable tasks, and review checkpoints. In a joint test reported by OpenAI and AWS, GPT-5.6 Terra completed successful Terminal-Bench 2.1 tasks at roughly 82% lower cost inside Kiro. That figure is a vendor test, not an independent guarantee for a customer codebase.

The more durable signal is distribution. OpenAI’s models now sit inside an AWS-native coding workflow alongside other options, giving teams more room to route simple and difficult work differently. Builders should compare completed, accepted changes per dollar and include review time, failed attempts, and regressions. A cheaper agent that creates a longer code-review queue has simply moved the cost into a meeting.

Google Cloud and Verizon push Gemini deeper into telecom operations

Google Cloud and Verizon announced a strategic partnership centered on Gemini Enterprise, data infrastructure, and AI-driven customer and network operations. The companies say Verizon will use Google’s stack to unify enterprise data, modernize customer experiences, equip employees with agents, and advance autonomous network operations. They did not disclose the agreement’s financial value or a detailed deployment timetable.

Telecom is a useful stress test for enterprise AI because errors meet real customers, regulated data, and physical networks. The value will come from measurable changes—shorter support resolution, fewer network incidents, or lower operating cost—not the number of agents deployed. Watch for production metrics, human escalation rules, and clarity on what data can flow into Gemini systems.

NVIDIA reportedly explores a Perplexity investment above a $30 billion valuation

NVIDIA is discussing an investment in Perplexity as part of a funding round that could value the AI search company above $30 billion, Reuters reported on August 24, citing The Information. Perplexity declined to comment, and no transaction has been announced. The valuation would be more than 50% above the company’s reported $20 billion valuation last year.

If completed, the deal would add another software and distribution bet to NVIDIA’s expanding portfolio. Perplexity gains capital and a closer relationship with the leading AI-chip supplier; NVIDIA gains exposure to a consumer and enterprise answer engine that drives inference demand. Until terms are confirmed, this belongs in the reported-plans column. Watch for the investor roster, deal size, governance rights, and whether the relationship includes compute commitments.

Watch & Learn

What is Mixture of Experts? — IBM Technology

IBM Technology gives a visual, approachable explanation of how a router activates only selected “experts” inside a large model. It is useful for product leaders and builders who see enormous parameter counts but want to understand actual compute. The video takes about eight minutes and requires no machine-learning background.

AI, Translated

Mixture of Experts (MoE)

A Mixture of Experts model contains several neural-network subcomponents called experts, plus a router that selects a small number for each token. The experts are not tidy job titles such as “lawyer” or “coder”; they learn different internal patterns. A model might contain 100 billion parameters but activate only 10 billion for a particular token. You should care because total model size alone can badly misrepresent inference speed, memory needs, and operating cost.

Try This Today

Create a cleaner AI-news visual with Gemini

Goal: produce a consistent 16:9 editorial image in about 10 minutes using Gemini’s Nano Banana 2 image model. Google says the model supports faster iteration, subject consistency, stronger text rendering, and production-ready aspect ratios. Availability and generation limits can vary by account, product, and region; Google’s announcement lists the supported surfaces.

1. In Gemini, choose image creation and describe the subject, composition, lighting, palette, and exclusions separately.
2. Generate one image first. Use follow-up edits for a single change at a time instead of rewriting the entire prompt.
3. Check logos, hands, numbers, and any visible text at full size before using the result.

Copy-ready prompt: “Create a sophisticated 16:9 editorial illustration for an AI business newsletter. Show a traditional research library transforming into a modern neural-network workspace, with analysts reviewing verified sources. Deep navy, warm amber, restrained cyan accents, clean negative space on the left for a headline added later. Photorealistic editorial style, natural lighting, no logos, no embedded text, no futuristic hologram clichés.”

One thing to remember

The strongest AI moat may be neither the largest model nor the biggest budget. It may be the combination of data, expertise, workflow, and proof that competitors cannot cheaply reproduce.