LastWeekIn.Tech

The biggest AI stories today center on economics and infrastructure—enterprise adoption of cheaper open-weight models, aggressive API price cuts, and massive capital raises for AI buildout—alongside a sobering capability check on medical AI and deepening US-China geopolitical fracture.

1. 1

   Number 1. Open Source Hacker News

   IBM CEO says open weight models are now good and cheap on-prem

   IBM CEO Arvind Krishna told investors on the company's Q3 earnings call that open-weight models now deliver roughly 90% of frontier model performance at about 10% of the cost, running on just two or three server racks on-premises. He contrasted this with 2024, when he said broad benchmark testing showed closed frontier models scoring around 70 versus 20-30 for open alternatives.

   Why it matters This is a significant enterprise validation signal: a major hyperscaler/enterprise vendor CEO is publicly endorsing open-weight models as production-ready substitutes for closed frontier systems in cost-sensitive, on-prem deployments. It suggests the performance gap that justified premium closed-model pricing is closing fast, with direct implications for enterprise AI budgets, data center sizing decisions, and the competitive position of closed-model providers like OpenAI and Anthropic.

   What to watch Watch for IBM's own product roadmap and whether other enterprise vendors echo this shift toward on-prem open-weight deployment in upcoming earnings calls.

2. 2

   Number 2. Industry & Economics Hacker News

   OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?

   OpenAI and Anthropic have cut LLM API pricing by up to 80% over the past several months amid escalating competition. The piece traces a timeline including OpenAI's GPT-5.6 Sol and Luna models and Anthropic's Claude Sonnet 5, each launched with progressively lower per-token input, output, and cache-read pricing.

   Why it matters Rapid, repeated price cuts across both leading closed-model providers point to intensifying margin pressure and commoditization at the mid-tier of the LLM market, likely driven by competition from cheaper open-weight and Chinese models. This reshapes the economics for startups and enterprises building on these APIs and raises questions about long-term profitability for frontier labs that depend on API revenue to fund massive compute investments.

   What to watch Watch whether further price cuts continue as open-weight and Chinese models undercut further, and how this pressure shows up in OpenAI's and Anthropic's financial disclosures.

3. 3

   Number 3. Capabilities Hacker News

   A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis

   A practicing dermatologist tested GPT-6 Astra and Claude Fable 5.1 against a decade-old convolutional neural network on skin cancer image diagnosis. Both frontier multimodal LLMs misclassified a melanoma as a benign seborrhoeic keratosis, while the older specialized CNN correctly flagged the cancer.

   Why it matters This is a concrete, domain-expert-verified data point showing that general-purpose frontier LLMs, despite sweeping benchmark gains, can still underperform narrow, purpose-built models in high-stakes visual diagnostic tasks. It's a useful counterweight to hype around AI matching or exceeding specialists in medicine, and a caution against deploying generalist models for clinical image reading without rigorous domain-specific validation.

   What to watch Expect more clinician-led benchmarking of frontier multimodal models against specialist tools before any push toward clinical deployment, and potential model-maker responses addressing diagnostic accuracy gaps.

4. 4

   Number 4. Infrastructure & Chips techmeme.com

   Meta reportedly plans to raise $95 billion to fund its AI expansion

   Meta is reportedly planning to raise approximately $95 billion to fund its AI infrastructure expansion, according to a Techmeme press-consensus report that also noted the US Treasury proposing new rules to restrict high-tech investment in AI and supercomputing by countries of concern.

   Why it matters A raise of this scale underscores just how capital-intensive the AI infrastructure race has become among hyperscalers, with Meta joining Microsoft, Google, and others in committing tens of billions toward data centers and compute. Combined with new Treasury restrictions on outbound high-tech investment, it highlights how financial and regulatory forces are simultaneously fueling and constraining the global AI buildout.

   What to watch Watch how Meta structures this financing (debt vs equity), where the capital is deployed, and whether the proposed Treasury investment-screening rule advances toward formal rulemaking.

5. 5

   Number 5. Policy & Regulation news4jax.com

   The U.S. and China deepen competing AI blocs

   Reports describe the US and China each building their own global AI blocs — a US-led grouping sometimes called 'Pax Silica' and a China-backed 'World AI Cooperation Organization' — with other countries increasingly forced to choose access to chips and AI technology from one side or the other as the window for neutrality narrows.

   Why it matters This formalizes a geopolitical bifurcation of the AI ecosystem that affects chip export control enforcement, international AI governance standard-setting, and the strategic calculus for countries caught between US and Chinese technology spheres. It has direct consequences for global supply chains, diplomatic alignment, and the pace at which AI capabilities diffuse unevenly across regions.

   What to watch Watch which countries formally align with each bloc in coming months and how this affects chip export licensing decisions and multilateral AI governance negotiations.