Weekly AI Radar: January 22-29, 2026
This has been one of the most intense weeks of the year: DeepSeek revolutionizes the market with R1, OpenAI and Anthropic compete for healthcare, and studies reveal that 95% of GenAI projects don’t generate measurable ROI. Meanwhile, Spain positions itself with an AI factory in Barcelona.
DeepSeek shakes the industry
The week’s dominant news didn’t come from Silicon Valley. DeepSeek R1, the Chinese open-source reasoning model, surpassed ChatGPT in downloads and caused a $600 billion drop in the semiconductor market in a single day.
Why R1 matters:
- Performance comparable to GPT-5 and Claude Opus 4.5 in reasoning and code
- Trained with a fraction of the budget (~$6M estimated vs hundreds of millions)
- Open-source with available weights - runs on 2x RTX 4090
- API with prices 90-95% lower than OpenAI
DeepSeek V4 is coming (estimated mid-February), with a 1M+ token context window and MoE architecture promising 50% additional computational cost reduction.
If you don’t know DeepSeek yet, I have a complete guide. And if you’re concerned about data sovereignty, I wrote about the implications of using Chinese models.
The healthcare war
Both companies launched products for the healthcare sector days apart. Healthcare is the enterprise battlefield of 2026.
OpenAI for Healthcare (Jan 16):
- HIPAA-compliant API
- Already deployed at Stanford Medicine, Memorial Sloan Kettering, UCSF
- Data point: 230 million users talk about health with ChatGPT weekly
Claude for Healthcare (Jan 12):
- Native connectors for medical databases: CMS, ICD-10, PubMed
- Focus on automating prior authorizations and clinical documentation
This fits the trend of vertical AIs I mentioned with Prism’s launch: the generic chatbot is giving way to domain-specialized tools.
Other product news
| Company | News | Why it matters |
|---|---|---|
| OpenAI | ChatGPT Go at $8/month | Accessible tier for emerging markets |
| OpenAI | Ads in free ChatGPT | Business model change |
| Gemini Personal Intelligence | Connects Gmail, Photos, YouTube - competing with Apple Intelligence | |
| Microsoft | Model selector in Copilot | ”Quick Response” vs “Think Deeper” |
| Anthropic | Claude Code reaches $1B ARR | 10x growth in enterprise integrations |
| Meta | Superintelligence Labs delivers models | Bosworth confirms results at Davos |
The uncomfortable stat: 95% without measurable ROI
This week studies were published confirming what many suspected: there’s an abyss between demos and production.
MIT Project NANDA
Analyzed 300 deployments and found only 5% of GenAI pilots generate P&L impact. Surviving projects have high domain specificity and deep integration into existing workflows.
Forrester State of AI Survey
- Only 13% report positive EBITDA impact
- 48% of companies already cut staff citing AI
- Only 14% commit to 3-year ROI horizons
Underestimated costs: 5-10x reality
Reports agree: companies underestimate AI project costs by 500% to 1000% when scaling from pilot to production. Typical breakdown: 30-40% technology, 60-70% implementation and change management.
This connects directly with what I wrote about the truth of implementing AI in SMBs. Real cost is always higher than budgeted.
Coding assistants: with nuance
The NAV IT study contradicts headlines: found no statistically significant changes in commit metrics after adopting Copilot. Active users were already active before the tool.
Additionally, 29.1% of Python code generated by Copilot contains security vulnerabilities.
Data tools
Microsoft acquires Osmos
Autonomous data engineering startup with agents. Integrates into Microsoft Fabric. Clear signal: data engineering is being automated.
Databricks Runtime 18.0 GA
Lakebase Autoscaling in preview (scale-to-zero, database branching, instant restore). Also launching Skills for Databricks Assistant following the Agent Skills standard.
BigQuery: GenAI functions now GA
AI.GENERATE, AI.GENERATE_TABLE, AI.EMBED, AI.SIMILARITY available. Allow analyzing text, images, video, and audio directly from SQL.
Google Data Catalog discontinued (Jan 30)
Action required: transition to Dataplex Catalog before the deadline.
For those working in data, this connects with GenBI and the future of analysts: tools are changing fast.
Regulation: Europe tightens, Spain prepares
AI Act: critical dates
- August 2, 2025 (in effect): Governance rules for GPAI models
- August 2, 2026: Obligations for high-risk systems (fines up to €35M or 7% revenue)
AESIA now operates with sanctioning power
Spain’s AI Supervision Agency has:
- Active inspection body since February 2025
- 16 technical guides published
- AI Sandbox with 12 high-risk projects completed
Spain gets an AI factory in Barcelona
EuroHPC selected Spain. The Barcelona Supercomputing Center (MareNostrum 5) will receive €197 million to develop an experimental AI model platform.
This is relevant to AI adoption in Spain I analyzed weeks ago.
Spanish ecosystem
2025 startup investment
- €3.108 billion total (-3% vs 2024, but +11% in number of operations)
- Barcelona leads with €1.374B, followed by Madrid (€865M)
- Mega-round: Multiverse Computing (San Sebastián) raised €189M
51% of Spanish startups already integrate AI
According to South Summit 2026, adoption went from 30% in 2024 to 51% now. There are 394 AI companies in Spain generating +5,000 jobs.
A concrete example of applied AI in Spain: the Data-ka case automating tender documents.
What few are covering
”AI Redundancy Washing”
Oxford Economics argues that companies are NOT replacing workers with AI at significant scale. Of recent tech layoffs:
- Only 4.5% cited AI as the reason
- 4x more lost jobs due to “market conditions”
Infrastructure: the invisible wall
More than 50% of AI projects have been delayed or canceled due to infrastructure complexity. The problem is no longer “can AI do this?” but “can we feed and cool it?”
OpenAI admits being “compute constrained”. This connects with the on-premise vs cloud debate: infrastructure matters more than ever.
AI agents as security threat
Palo Alto Networks warns that 80% of companies deploy agents without proper governance. Risks: agents with broad permissions, unsolved prompt injection, documented cyberattack automation.
Emerging trends
Small Language Models
Gartner predicts that by 2027, organizations will use SLMs 3x more than general LLMs. Phi-3.5 (3.8B parameters) rivals models 10x larger using 98% less computational power.
MCP consolidates
Anthropic’s Model Context Protocol establishes itself as a standard. Gartner estimates 40% of enterprise apps will integrate agents by end of 2026.
Key dates
| Date | Event |
|---|---|
| January 30, 2026 | Google Data Catalog discontinued |
| February 1, 2026 | BigQuery: multi-region transfer charges |
| Mid-February 2026 | DeepSeek V4 (estimated) |
| August 2, 2026 | AI Act: high-risk obligations |
Conclusion
2026 won’t be the year of AI collapse, but of necessary correction. Data is clear: 95% of pilots don’t scale, real costs exceed estimates by 5x, and infrastructure is the new bottleneck.
Winning companies will be those that abandon “hype mode” and focus on specific implementations with demonstrable ROI. DeepSeek proved the frontier doesn’t require billions, just intelligent architecture.
Next week: watch for DeepSeek V4 and OpenAI/Anthropic reactions.
Which news seemed most relevant to you? Am I missing something important?
You might also like
AI in Spain 2025: adoption, barriers, and real opportunities
Complete analysis of AI in Spain: adoption data, global comparison, the three big barriers, and opportunity sectors. Updated January 2026.
OpenAI launches Prism: the end of the generic chatbot has begun
A free workspace with GPT-5.2 for writing papers in LaTeX. But the real news isn't Prism: it's what it represents about where AI is heading.
The AI bubble: 7 trillion looking for returns
Who wins, who loses, and why you should care. Analysis of massive AI investment and its bubble signals.