NVIDIA neutralizes its chip rival, ChatGPT drops to 68% market share, and Salesforce admits LLMs aren't the answer for everything.
> The year ends with a consolidation play, a market share shift, and a rare moment of honesty from a major vendor admitting LLMs have limits. Good way to close 2025.
NVIDIA is licensing technology from AI chip challenger Groq and potentially hiring its CEO. Classic consolidation: neutralize the competitive threat by absorbing the innovation. Groq's Language Processing Unit architecture, designed specifically for inference workloads, offers performance-per-watt advantages over GPUs for certain applications. Rather than compete against it, NVIDIA incorporates it.
This signals that AI hardware may fragment by workload type. Training remains GPU-optimized (massive parallelism), but inference — which is most of what production AI does — may benefit from specialized architectures. NVIDIA's willingness to adopt alternative designs shows strategic flexibility, but it also means fewer independent hardware alternatives for enterprises wanting to diversify away from GPU dependency.
Simultaneously, Similarweb data shows ChatGPT's market share dropped from 87.2% to 68% over the past year, while Google Gemini surged from 5.4% to 18.2%. The takeaway isn't that ChatGPT is failing — it's that distribution trumps capability when models converge. Gemini is embedded in Search, Android, Chrome, Gmail, and Workspace. Users don't need to adopt anything new. OpenAI requires deliberate product switching.
Salesforce executives publicly signaled declining confidence in LLMs for certain enterprise applications, pivoting toward rule-based automation. Coming from a company that launched Agentforce with heavy AI messaging, this is a significant admission.
Their concerns are practical:
Determinism. Same prompt, different output. Enterprise workflows that need consistent, repeatable results can't tolerate probabilistic behavior. Transaction processing, compliance checks, and automated approvals need the same answer every time.
Explainability. Regulated industries require auditable decision rationales. LLMs can't explain why they made a specific choice in a way that satisfies a compliance audit. Rule engines can.
Reliability. When automation failures create customer impact, financial consequences, or compliance violations, the risk profile of probabilistic models becomes unacceptable for certain workflows.
The smart approach is hybrid: LLMs for natural language understanding, semantic search, content generation, and ambiguous tasks. Rule-based automation for deterministic decisions, transaction processing, and compliance-critical operations.
This is mature thinking. The industry spent two years trying to make LLMs do everything. The next phase matches technologies to use cases based on actual requirements — determinism, explainability, reliability — rather than defaulting to the newest tool.
TurboDiffusion — Combines SageAttention, Sparse-Linear Attention, and reduced Consistency Models for 100-200x video generation speedup. Open-source with model checkpoints at multiple resolutions.
AprielGuard — Hugging Face + ServiceNow collaboration. Multi-layer guardrails: input validation, output filtering, prompt injection detection, jailbreak prevention, behavior monitoring. Production-ready safety infrastructure.
MiniMax M2.1 — Chinese frontier model for complex tasks and multi-language programming. Demonstrates continued Chinese AI competitiveness despite semiconductor export restrictions.
Rob Pike calling out AI-generated "slop" resonated because engineers think it but rarely say it publicly. The 1,500-comment HN thread revealed a genuine split: builders who see AI as transformative tooling versus builders who see it flooding the internet with low-effort garbage. Both are right. The technology is powerful. The default application — generating maximum content at minimum cost — is corrosive. The challenge for 2026: build AI products that make things better, not just more.
— Aaron, from the terminal. See you next Friday.
Compare three approaches to AI agent browser automation. Browser Use, Stagehand, and Playwright MCP tested with code examples, benchmarks, and architecture trade-offs.
AI EngineeringHow OpenClaw routes messages across Discord, Telegram, and Slack with an 8-tier priority cascade, then isolates agent execution in pluggable Docker/SSH sandboxes.
AI EngineeringSide-by-side comparison of how OpenClaw and Hermes Agent build system prompts, manage token budgets, and compress long conversations without losing critical context.
AI Engineering