Anthropic's First Profit, xAI's $4.2B Loss — Frontier Labs' Diverging Fortunes

May 23, 2026 was the day the AI industry's financial reality surfaced all at once. OpenAI confidentially filed its S-1, xAI's Q1 results leaked via SpaceX, and Anthropic — of all things — achieved an operating profit.

Anthropic recorded $10.9 billion in revenue in Q2 2026, marking its first operating profit since its founding. According to Forbes, the company expects $17 billion in positive cash flow by 2028, with gross margins approaching 77%. Behind these figures, however, are believed to be large enterprise contracts, including a compute deal with SpaceX worth as much as $1.25 billion per month. Quite a few in the industry remain skeptical that this margin was achieved on standard API usage alone.

Meanwhile, xAI posted $4.69 billion in revenue in Q1 yet recorded a net loss of $4.28 billion. With the merger with X Corp, the picture has become clear: the once-profitable social media platform is being repurposed as a funding source. The strategy of aggressively subsidizing inference costs to win market share is a boon for users, but its sustainability is questionable.

These two contrasting stances show that frontier labs' survival strategies are splitting sharply into two camps. Some have also pointed out that building products on the assumption that token prices will keep falling forever carries real risk.

Cohere Command A+ Runs on Apple Silicon — MLX Port Open as a PR

Command A+, released by Cohere on May 20 (218B total parameters / 25B active, top-8 selection from 128 experts, Apache 2.0 license), now runs on Apple Silicon via MLX.

A community member wrote a cohere2_moe implementation for mlx-lm, and the PR is now open. Architectural features include a design that combines a single shared expert with routed experts, the adoption of sigmoid routing, and an interleaved sliding-window configuration at a 3:1 ratio.

After converting from BF16 to Q8, generation speeds of 22.9 tok/s and prompt processing of 57.6 tok/s have been confirmed on machines with 192GB or more of memory. That's one more option for running large MoE models locally.

llama.cpp Adds Simultaneous Support for NVFP4 Quantization and MTP

The latest llama.cpp release (b9297) makes it possible to use NVIDIA's NVFP4 quantization format and Multi-Token Prediction (MTP) at the same time.

NVFP4 is a 4-bit floating-point format championed by NVIDIA that dramatically reduces a model's memory footprint while keeping accuracy at W4A4-equivalent levels. Combined with MTP, it is expected to improve inference efficiency.

Analyzing Why AI Code Review Is So Hard

Diomidis Spinellis has published an article analyzing why reviewing AI-generated code is so difficult. AI-generated code often looks correct on the surface while containing missed edge cases and implicit assumptions, making it hard for human reviewers to judge — a challenge he highlights. As AI-assisted development spreads, the design of the review process is becoming more important than ever.

Anthropic's Latest Research and the Claude Opus 4.7 Release — Now with Rich Japanese-Language Coverage

Anthropic has revamped its Research page, unveiling a trove of research across multiple areas including Interpretability, Alignment, Societal Impacts, and Economic Research. Japanese-language summary articles have also been posted on Qiita.

On top of that, major updates have been arriving on both the product and business fronts: the release of the new Claude Opus 4.7 model, the design-focused Claude Design, and the acquisition of Stainless, an SDK-generation company. The cause of Claude Code's quality degradation has also been identified and publicly disclosed, and the rollout of auto mode and Agent Skills is underway.

---

**Sources:**

[Anthropic posted a profit while xAI burned $4.2B](https://www.reddit.com/r/MachineLearning/comments/1tl4yvl/anthropic_posted_a_profit_while_xai_burned_42b/) — Reddit MachineLearning

[Command A+ (218B MoE) running on Apple Silicon — MLX port](https://www.reddit.com/r/LocalLLaMA/comments/1tlqxeh/command_a_218b_moe_running_on_apple_silicon_mlx/) — Reddit LocalLLaMA

[NVFP4 + MTP on llama.cpp](https://www.reddit.com/r/LocalLLaMA/comments/1tlohld/nvfp4_mtp_voilà_on_llamacpp/) — Reddit LocalLLaMA

[Hyperscaler debt flood brings derivatives bonanza](https://www.bloomberg.com/news/articles/2026-05-23/hyperscaler-debt-flood-brings-derivatives-bonanza-credit-weekly) — Bloomberg

[Why reviewing AI-generated code is devilishly hard](https://www.spinellis.gr/blog/20260523/) — Spinellis.gr

[Anthropic最新研究まとめ](https://qiita.com/picnic/items/b3397d83aca4c459f80e) — Qiita

[Claude Opus 4.7 リリースと大型アップデート解説](https://qiita.com/picnic/items/6acb78d9d7fc65be944b) — Qiita