Rakuten Group made quite the entrance on March 17, 2026. The Japanese e-commerce giant unveiled Rakuten AI 3.0 with serious fanfare — branding it "Japan's largest and most powerful" AI model, a roughly 700-billion-parameter Mixture-of-Experts system built with government cash from Japan's Ministry of Economy, Trade and Industry. Chief AI Officer Ting Cai gushed about an "outstanding combination of data, engineering, and innovative architecture at scale." The press release? Not a single word about any third-party foundation model. Within hours, the open-source community had torn the whole story apart.
The Config File That Blew Everything Up
No whistleblower was needed. Just a config file. Developers poking around Rakuten's Hugging Face repository spotted "model_type": "deepseek_v3" sitting right there in the configuration JSON — paired with a parameter structure that matched DeepSeek-V3 almost exactly, down to its 671-billion total and 37-billion activated parameters. That's the Chinese open-source model dropped in late 2024. Suddenly, Rakuten's slippery PR line about incorporating "the best of the open-source community" looked less like modesty and more like a deliberate smokescreen. The architecture signatures, parameter counts, and config details weren't just similar — they were forensically identical. This wasn't inspiration. It was derivation, plain and simple.
One Rule. One Violation.
Here's the almost poetic irony at the center of this mess. DeepSeek-V3 runs under the MIT license — arguably the most generous license in all of software. MIT asks for exactly one thing in return: keep the original copyright notice. That's it. Rakuten didn't. The company quietly deleted the DeepSeek license file and swapped in an Apache 2.0 license under its own name. Why? Apache 2.0 carries heavier enterprise credibility and makes a release smell more like original authorship. The move backfired spectacularly. Deleting that one notice wasn't a grey area or a technicality — it was a concrete, enforceable license breach.
After the community surfaced the evidence, Rakuten restored an acknowledgment file. Critics, though, pointed out the fix landed under a filename called "NOTICE" rather than restoring the actual license — a reactive patch job, not a genuine correction.
This Goes Way Beyond Rakuten
Smart investors shouldn't write this off as one company's embarrassing stumble. It's a flare going up over a structural problem spreading across the entire AI industry. The 2026 OSSRA report found that 68% of audited codebases now carry open-source license conflicts — an all-time high. Much of that spike traces directly to what researchers call "license laundering," where AI-derived code loses its upstream attribution as it passes through fine-tuning pipelines. Worse, only 54% of organizations currently check AI-generated or AI-derived code for IP and licensing risks. The rest are quietly stacking legal debt that surfaces at the absolute worst moments: product launches, M&A due diligence, regulatory audits.
The government angle makes this murkier still. Japan's METI backed the launch without apparently asking where the model actually came from — a procurement blind spot that regulators worldwide will study closely. The EU AI Act already demands training data transparency for general-purpose AI models and tougher disclosure norms are spreading fast. The era when "we used open source" counted as a full answer is closing.
Provenance Is Now a Core Asset
The hardest lesson here cuts deepest for capital allocators and technology executives. Fine-tuning someone else's base model and presenting the result as proprietary technology is both common practice and increasingly risky. The open-source community now wields forensic tools — config files, architecture signatures, parameter counts — that make provenance claims checkable by anyone with a GitHub account. Overclaiming origin today means disproportionate reputational damage for a marginal branding gain. The math simply doesn't work anymore.
The fix isn't complicated. AI teams need a Model Bill of Materials — a documented chain of custody covering every base model, dataset, and fine-tuning asset, with license terms attached, similar to the software SBOM already standard in DevSecOps. Automated license scanning in CI/CD pipelines catches stripping or conflicts before public release. Legal sign-off before publishing any model built on third-party weights isn't optional overhead anymore. It's table stakes.
For anyone evaluating AI-native companies, model provenance documentation belongs right next to financial statements in due diligence. Can the company show a clean chain of custody from base model to product? If not, the liability is already there — just waiting to be discovered.
The cover-up, as always, ends up far worse than the foundation.
not investment advice