Data:October 19, 2025
Category:AI Models & Hardware
Reading time:12 minutes
Introduction
The last days of October 2025 marked a turning point in the artificial intelligence industry. Amid billion-dollar deals, technological breakthroughs, and bold financial projections, a complex picture emerges that deserves in-depth analysis. This article examines the main developments and offers a critical perspective on the implications of these advancements.
The Main News
OpenAI DevDay 2025: New Tools for Developers
On October 6-7, OpenAI held its largest DevDay ever, with over 1,500 attendees and significant announcements for the developer community.
AgentKitrepresents the complete framework for building, deploying, and optimizing agent workflows. During a live demo, an engineer from OpenAI created and published a functioning agent in less than 8 minutes, demonstrating the speed of the new system.
Apps SDKmarks a paradigm shift: developers can now build complete applications within ChatGPT, not just simple GPTs. The integration with Anthropic's Model Context Protocol (MCP) represents a step towards interoperability.
Codex, the coding agent from OpenAI, has exited the preview phase and is now available to all users. Usage has increased tenfold since the beginning of August 2025.
Finally,GPT-5 Pro and Sora 2have been made available via API, allowing developers to integrate advanced reasoning and video generation capabilities into their products.
The OpenAI-AMD Agreement: The Chip War
The most disruptive announcement of the month was the multi-billion dollar agreement between OpenAI and AMD for the deployment of 6 gigawatts of AMD chips in future AI data centers.
The contract provides for the supply of AMD Instinct MI450 GPUs starting in 2026, with OpenAI receiving warrants for up to 10% of AMD's shares. AMD shares rose by 38% after the announcement, and Barclays analysts project over $100 billion in revenue from the agreement over the next five years.
This deal represents the first real challenge to NVIDIA's dominance in the AI chip market, signaling the beginning of an infrastructural "arms race" where access to chips becomes the true barrier to entry for AI innovation.
Anthropic: Stratospheric Revenue Ambitions
Anthropic, the company behind Claude, has announced bold financial projections: from $7 billion in annualized revenue today to $20-26 billion by 2026, nearly tripling revenue in just over a year.
Anthropic's strategy differs from OpenAI: 80% of its revenue comes from enterprise clients, with significantly higher prices. Claude Opus 4 costs $15 per million input tokens and $75 per million output tokens, 5-7 times more than GPT-5.
The partnership with IBM, announced this month, aims to integrate Claude into IBM's development tools, focusing on enterprise security and automation. Internal tests have shown productivity gains of 45% in software testing tasks.
Samsung TRM: The Paradox of Efficiency
In contrast to the trend towards gigantism, Samsung has introduced the Tiny Recursion Model (TRM), a model with only 7 million parameters that outperforms billion-parameter LLMs in complex reasoning tasks.
TRM achieves 44.6% accuracy on ARC-AGI-1 and 7.8% on ARC-AGI-2, surpassing Gemini 2.5 Pro (4.9%) despite being 1000 times smaller. On Sudoku-Extreme, TRM obtains 87.4% accuracy, a huge leap from the 56.5% of its predecessor.
The recursive architecture allows the model to iteratively self-correct, demonstrating that algorithmic intelligence can surpass the brute force of scaling.
Google Veo 3.1: Evolution of Video Generation
Google has released Veo 3.1, an update to its video generation model that introduces improved audio, granular editing controls, and better output for image-to-video conversion.
The model is integrated into Flow, Google's AI video editor, which has generated over 275 million videos since May 2025. Veo 3.1 allows users to add objects to videos and blend them into the style of the clip, with the future possibility of removing existing objects.
European Union: AI Sovereignty Strategy
The EU has unveiled its "Apply AI Strategy" with an investment of 1 billion euros to reduce technological dependence on the USA and China. The focus is on defense, healthcare, and public administration, with support for European startups such as Mistral and Helsing.
The strategy embraces open-source technologies to foster the growth of startups and aims to develop defense command systems enhanced by AI.
Critical Analysis
The Infrastructure Arms Race
The OpenAI-AMD agreement confirms a worrying trend: AI is becoming a "chip war" rather than a competition of algorithmic innovation.
Positive aspects:The breaking of NVIDIA's monopoly creates healthy competition and could lower costs in the long term.
Issues:If access to chips becomes the true barrier to entry, we are consolidating power in the hands of a few tech giants who can afford billion-dollar investments. Moreover, 6 gigawatts of energy consumption is equivalent to the needs of a small nation, raising enormous environmental issues that are systematically ignored in announcements.
This deal demonstrates that the industry is betting everything on "brute force scaling" rather than efficiency. Is this a sustainable approach?
The Samsung TRM Paradox
Samsung TRM exposes a fundamental contradiction in the AI industry: a model 1000 times smaller can outperform giants in complex reasoning. What does this say about the direction of the industry?
Revolutionary Implications:TRM demonstrates that intelligent architecture surpasses raw dimensions, with a drastic reduction in computational costs and environmental impact.
Limitations:The model excels in structured reasoning (puzzles, logic) but not in natural language, suggesting that different approaches serve different objectives.
The critical question:If such a small model can achieve these results, is the industry wasting enormous resources chasing parameters instead of intelligence? While Samsung demonstrates efficiency, the industry invests trillions in scaling. This paradox should concern us.
The Revenue Bubble of Anthropic
Anthropic's projections raised serious doubts about the sustainability of the AI business model.
Red flag number one:A valuation of 183 billion with less than 10 billion in revenue and billion-dollar losses. OpenAI, valued at 500 billion, could generate only 13 billion in 2025 and lost 5 billion last year.
Red flag number two:Prices 5-7 times higher than OpenAI. How sustainable is it when competition lowers prices?
Red flag number three:According to a study by MIT, 95% of enterprise AI implementations have no measurable impact on profits and losses. How can Anthropic triple its revenues when companies are discovering that AI does not increase productivity as promised?
The consumer market is illusory:Only 5% of ChatGPT users pay for the service. Anthropic's enterprise focus is smart, but enterprise AI adoption is slowing down, not accelerating.
These projections seem more like marketing for investors than realistic business plans.
OpenAI AgentKit: Democratization or Lock-in?
The opening of the OpenAI platform to developers raises strategic questions.
Positive signals:The adoption of Anthropic's Model Context Protocol (MCP) is a good sign towards interoperability.
Legitimate doubts:Are we witnessing true democratization or the creation of a closed ecosystem like the App Store? Does monetization for developers imply that OpenAI takes a percentage?
The pattern resembles Apple's strategy: open the platform but control the marketplace. OpenAI is building a walled garden where it maintains ultimate control.
The EU Strategy: Too Little, Too Late?
The European investment of 1 billion euros, although necessary for technological sovereignty, seems inadequate.
The scale problem:1 billion is nothing compared to the 100+ billion of the OpenAI-AMD deal, the 13 billion from Microsoft in OpenAI, or the 10+ billion funding of Anthropic.
Fragmentation:Europe has talent (Mistral, Helsing) but lacks coordination and scale. The investment comes as the USA and China already have a 5+ year advantage.
Concrete risk:Europe risks remaining a consumer of technology rather than a producer, with significant geopolitical implications.
The Ignored Energy Problem
No announcement mentions the environmental cost, yet it is critical.
The numbers:6 gigawatts for OpenAI-AMD equate to the consumption of a small nation. The training of GPT-5 generates emissions equivalent to thousands of transatlantic flights.
The contradiction:Big Tech promises sustainability but invests in energy-intensive AI. Samsung TRM shows that efficient alternatives exist, but the industry ignores them because scaling means easier marketing.
The AI industry is creating an energy crisis while everyone celebrates the advancements. This silence is deafening.
Trends and Forecasts
Worrying Trends
Financial bubble:Astronomical valuations without real profits suggest an imminent correction.
Infrastructure arms race:The competition is shifting from models to chips, favoring only the tech giants.
Greenwashing:The environmental impact is systematically ignored in advertisements.
Consolidation:Only a few players can afford to compete, reducing innovation.
Hype vs reality:The promises of enterprise productivity are not materializing as expected.
Positive Signals
Proven efficiency:Samsung TRM proves that sustainable alternatives exist.
Open standards:MCP is a step towards interoperability.
Hardware competition:AMD vs NVIDIA could lower costs.
Specialization:Anthropic on enterprise, OpenAI on consumer creates diversification.
Forecasts for the Next 6-12 Months
Market correction:Current valuations are unsustainable and a correction is likely.
Energy crisis:Let's expect regulations on the energy consumption of AI.
Disillusion enterprise:More companies will discover that the ROI of AI is overstated.
Europe behind:The technological gap with the USA and China will widen.
Shift towards efficiency:More research on small and efficient models like TRM.
Conclusion: The Fundamental Question
Are we building artificial intelligence or are we just building bigger data centers?
Samsung TRM suggests that the answer might be the second one. While the industry celebrates billion-dollar deals and stratospheric revenue projections, a model with 7 million parameters demonstrates that intelligence does not necessarily require brute force.
The implications are profound:
Sustainability:The current model is unsustainable from an energy and financial perspective.
Accessibility:If only tech giants can compete, democratic innovation dies.
Effectiveness:The promises of enterprise transformation are not materializing.
Direction:The industry is optimizing for marketing and valuations, not for real intelligence.
October 2025 may be remembered not for the announced breakthroughs, but as the moment when the contradictions of the AI industry became impossible to ignore. The question is not whether there will be a correction, but when and how painful it will be.
Artificial intelligence has enormous potential, but only if the industry changes direction: from scaling to efficiency, from hype to measurable results, from unlimited energy consumption to real sustainability.
Samsung TRM shows the way. It remains to be seen if the industry will have the courage to follow it.
---
Sources:
- OpenAI DevDay 2025 (ZDNET)
- Medium: "October 2025: Major AI Developments"
- TechCrunch: Google Veo 3.1 Release
- AI News: Samsung TRM Research
- Tom's Hardware: Anthropic Revenue Projections
- Various industry reports and analyses