As enterprise and solution architects, we’re tasked with making strategic technology decisions that will serve our organisations for years to come. But what happens when the technology we’re being asked to implement is built on fundamentally unsustainable economics? The 2025 financial data from major AI providers presents a sobering reality check that every architect needs to understand.
Almost exactly a year ago, I warned in Will generative AI implode and become more sustainable? that brute-force approaches to AI development were fundamentally unsustainable. Twelve months later, the numbers tell a story that should concern any architect responsible for long-term technology strategy. OpenAI has reached $10 billion annualised revenue but loses approximately $5 billion annually. Anthropic hit $4 billion in revenue while expecting to lose $3 billion in 2025. This isn’t just poor financial management – it’s what tech critic Ed Zitron calls the “subprime AI crisis,” drawing parallels to 2007’s financial meltdown.
For architects, this isn’t just an interesting financial footnote. It’s a fundamental risk to any AI strategy built around these providers.
The Architecture Risk Assessment
When we evaluate technology vendors, financial stability typically features prominently in our risk assessments. Yet somehow, with AI providers, we’ve been willing to overlook burn rates that would disqualify any traditional software vendor.
Consider the numbers from an architectural perspective:
- OpenAI: 50% loss ratio on revenue
- Anthropic: 75% loss ratio expected for 2025
- Combined Big Tech AI capex: $250 billion for 2025
- Sector venture funding: $100+ billion consumed in 2024
These aren’t the metrics of a mature, stable technology platform. They’re the metrics of an experimental industry hoping scale will eventually solve fundamental economic problems.
Architectural Implication: Any enterprise architecture built around these providers carries significant continuity risk. When your AI strategy depends on companies burning billions annually with no clear path to profitability, you’re essentially betting your organisation’s future on venture capital continuing indefinitely.
The Vendor Lock-in Trap
The current AI landscape presents a potentially dangerous form of vendor lock-in. Unlike traditional software where you might struggle to migrate data or retrain users, AI model dependencies create deeper architectural coupling:
- Model-specific prompt engineering: Optimisations for GPT-4 don’t necessarily transfer to Claude or other models, requiring extensive re-testing and tuning
- Performance assumptions: Applications designed around current subsidised pricing and performance levels
- Training data lock-in: Fine-tuned models tied to specific provider platforms
- Integration complexity: While most providers have adopted OpenAI’s API format as a de facto standard, making switching technically easier, the real challenge lies in ensuring prompts and workflows perform consistently across different models
Architectural Recommendation: Design for provider agnosticism from day one. Test critical prompts across multiple providers during development. This isn’t just good architectural practice – it’s essential business continuity planning.
The Open Source Alternative Architecture
The 46% enterprise preference shift towards open source models isn’t just about cost – it’s about architectural control and risk mitigation. Open source models offer several architectural advantages:
Deployment Flexibility: Host locally, in your cloud, or hybrid arrangements based on your requirements rather than provider limitations.
Version Control: Pin to specific model versions without forced upgrades that might break your applications.
Compliance Certainty: Full control over data processing, storage, and geographic restrictions.
Cost Predictability: Hardware costs are predictable; API costs from financially unstable providers are not.
Performance Tuning: Ability to optimise models for your specific use cases without depending on general-purpose commercial models.
Examples of this shift include ByteDance’s Doubao at $0.0001 per 1,000 tokens – 99.8% cheaper than GPT-4 pricing, and Alibaba’s Qwen models pricing 83% below Western equivalents. While these Chinese models benefit from $912 billion in government funding, they also demonstrate the efficiency gains possible through smaller, more focused models.
Architectural Pattern: Consider a “hybrid AI architecture” where critical functions use locally-hosted open source models for reliability, while non-critical applications can leverage commercial APIs for convenience. This provides the best of both worlds while managing risk.
Energy and Infrastructure Architecture Considerations
Sam Altman’s recent disclosure that each ChatGPT query uses 0.34 watt-hours might sound minimal, but the architectural implications are significant. At 1 billion queries per day, that’s 340 megawatt-hours daily for a single service. Research suggests ChatGPT already uses nearly 40 million kilowatts of energy per day – enough to power substantial regions.
For enterprise architects, this translates to:
Infrastructure Planning: AI workloads require fundamentally different infrastructure considerations than traditional applications. Data centre power and cooling requirements are an order of magnitude higher than traditional compute workloads, creating significant capacity planning challenges.
Cost Modelling: Current AI pricing is heavily subsidised. When providers need to reach profitability, enterprise architects should expect significant price increases – potentially 5-10x current levels based on the loss ratios we’re seeing.
Sustainability Governance: If your organisation has net-zero commitments, AI implementations need careful energy accounting. Hugging Face’s AI Energy Score project provides a framework for measuring and comparing model efficiency – essential for architects balancing functionality with environmental responsibilities.
Practical Implementation Strategies
Near-term (6-12 months):
- Audit existing AI dependencies and assess provider risk
- Implement abstraction layers around AI API calls
- Pilot open source alternatives for non-critical applications
- Establish energy/cost monitoring for AI workloads
Medium to Long-term (1-5 years):
- Develop hybrid architectures combining local and cloud-based AI
- Invest in team capabilities for open source model deployment and management
- Build core AI capabilities around sustainable, controllable models
- Create procurement frameworks that include sustainability and financial stability criteria
- Develop neurosymbolic approaches that combine AI with deterministic logic
- Plan for significant cost increases from commercial AI providers
- Create vendor-agnostic AI platforms that can adapt to market changes
- Establish AI governance frameworks that prioritise efficiency over scale
Striving for Efficient AI Architecture Patterns
- Right-size the model: Use the smallest model that achieves acceptable results
- Hybrid processing: Combine AI with traditional algorithms where appropriate
- Edge deployment: Process locally where possible to reduce latency and costs
- Caching and reuse: Avoid redundant AI calls through intelligent caching
- Fallback strategies: Design graceful degradation when AI services are unavailable
These approaches align with both cost management and sustainability goals while reducing dependency on financially unstable providers.
The Governance Imperative
Architecture review boards need to start asking harder questions about AI implementations:
- What’s our continuity plan if this AI provider significantly raises prices or goes under?
- How does this AI implementation align with our sustainability commitments?
- Are we building capabilities or just dependencies?
- What’s the total cost of ownership including energy, infrastructure, and risk mitigation?
- How do we measure and optimise the efficiency of our AI implementations?
The transparency initiatives from organisations like Hugging Face provide frameworks for these assessments, but governance processes need to evolve to incorporate sustainability and financial stability as first-class concerns. As I discussed recently in my analysis of Enterprise AI deployment patterns, the gap between executive expectations and actual value creation (75% of executives naming AI as a top-three priority while only 25% report creating significant value) represents a fundamental architecture challenge.
Looking Forward: The Sustainable AI Architecture
The market correction I predicted appears increasingly likely, with 2026-2027 as the critical period when unsustainable approaches face reality. Training costs for next-generation models approach $1 billion, whilst data centre requirements expand from 50-200 megawatts to over 1 gigawatt. Average AI engineer compensation reaches $925,000 at leading companies – creating exactly the unsustainable resource consumption trajectory I warned about in my original sustainability analysis.
Architects who prepare now will be best positioned to navigate this transition.
The future belongs to organisations that build AI capabilities rather than just AI dependencies. This means investing in understanding open source models, developing internal AI expertise, and creating architectures that can adapt to a changing provider landscape.
As the financial sustainability of major AI providers becomes questionable, the enterprises that thrive will be those with architectures designed for resilience, efficiency, and independence. The current “subprime AI crisis” isn’t just a financial problem – it’s an architectural opportunity for those willing to think beyond the hype and build for the long term.
The question isn’t whether the current AI economics are sustainable – they clearly aren’t. The question is whether your architecture will survive the inevitable correction.
Footnote – this article was reworked for an architecture audience from: GenAI sustainability: a review of the 2025 numbers from the Scott Logic blog.