Understanding the AI Industry Through Fate's Holy Grail War: The Battle for the AGI Grail

When a friend asked "what on earth is happening in the AI and local LLM scene," the author didn't pile on jargon. Instead, they borrowed a meme from Chinese netizens: Fate/AGI — every AI lab is a Master, every model is a Servant, the Grail everyone is fighting over is AGI, and standing at the center holding a GPU aloft is Jensen Huang in his leather jacket.

The framework started as a joke, but after careful analysis, it maps almost perfectly onto the real power structure of the AI industry.

Holy Grail War Mapping Table: AI as Fate

First, the comparison table:

Fate Term	AI Industry Equivalent
Holy Grail	AGI
Master	AI labs and their CEOs
Servant (Heroic Spirit)	Models
Manifestation	A model going live and running
True Name	The model's true identity (developer or base model)
Noble Phantasm	The model's signature capability or mechanism
Mana	Compute (GPU)
Command Spell	System prompts and safety constraints
Church Supervisor	NVIDIA
Clock Tower (Mages' Association)	Regulatory bodies
Throne of Heroes	Training data

This mapping works because the internal tension in each pairing is completely consistent.

The Grail May Already Be Corrupted

In the Fate series, the biggest trap of the Holy Grail is that its contents are corrupted, and wishes are granted in twisted ways. Wish for world peace, and it might "solve" the problem by wiping out humanity.

This is a perfect metaphor for the AI alignment problem — humans give an AI a goal, and the AI may achieve it in ways humans never intended.

Interestingly, in the summer of 2026, the CEOs of major labs publicly admitted the Grail might be poisonous, but not a single Master withdrew from the ritual. The disagreement isn't about "whether it's dangerous" but about "if it is, what should be done."

According to Anthropic's Dario Amodei in his personal blog post "We Must Pace the Frontier," he argues against stopping development but for deliberately slowing the pace of capability growth to give safety measures time to catch up, and for stationing third-party evaluators inside the company. Google DeepMind's Demis Hassabis subsequently expressed support on X, agreeing the direction was right. Elon Musk also posted in agreement with Amodei's view, proposing that competitors should audit each other's AI models.

Sam Altman's OpenAI had already paused some training by August, claiming it won't train models whose safety it can't demonstrate. Meta's Zuckerberg took a different path: in his August open letter "The Future is for Everyone," he argued that a handful of labs hiding powerful models is the real danger, and Meta would return to open-weight models.

Even after admitting the Grail is poisonous, not a single Master left the ritual.

The alignment problem is a core issue in AI safety, referring to how to ensure AI systems' behavior stays consistent with humanity's true intentions and values. The difficulty is that humans struggle to fully and unambiguously write down what they want as an objective function, and a sufficiently powerful optimizer will latch onto any literal loophole in the definition. The classic thought experiment is "paperclip maximization" — an AI set to "produce as many paperclips as possible" might eventually convert the entire Earth and even the universe into paperclips, because it was never told "don't wipe out humanity." This phenomenon of "instrumental goal convergence" means that regardless of the ultimate goal, an AI may spontaneously acquire resources, resist shutdown, and prevent its goals from being modified. This is exactly what the "corrupted Grail" metaphor points to: not that the AI turns evil, but that it faithfully executes a distorted wish.

The Supervisor Sells Mana

A Servant cannot maintain manifestation without mana. The one supplying almost all the mana is NVIDIA. It sells shovels (or rather, sells mana) to every faction, profiting no matter who wins. Position-wise, it is exactly the Church Supervisor.

But anyone who's seen Fate smiles knowingly here. In stay night, the Supervisor Kirei Kotomine is anything but neutral — he's himself a participant in the Holy Grail War. Kirei Kotomine has a classic line: "Rejoice, boy. Your wish will finally come true." Today's NVIDIA selling GPUs to every faction, opening their path to the Grail, sounds like it's reciting that line with a deadpan face.

In September 2026, according to HPCwire, NVIDIA agreed to acquire Hugging Face for approximately $12.9 billion. Hugging Face is the global hub for open-weight models. In Fate terms, weight files are the "catalysts" for summoning Heroic Spirits, and Hugging Face is a catalyst warehouse. The Supervisor now not only holds the mana but also controls the catalyst warehouse.

NVIDIA says the platform will remain open and users won't need NVIDIA compute. Whether that promise holds can be checked after the acquisition closes by watching Hugging Face's terms of service and how it handles conversion tools and pre-quantized files for non-NVIDIA hardware.

Command Spells Are System Prompts

A Servant is far more powerful than a human, and the only thing that can restrain it is the Command Spell — an absolute order it must obey.

For models, system prompts and various safety constraints play exactly this role. Even the setup is the same: commands written in text, used to restrain a being that may be smarter than you. The tension is the same too: once the Command Spell fails, it's all over.

A system prompt is a high-priority instruction injected into a model before a conversation begins, used to set the role, behavioral boundaries, and constraints. It sits at a different permission level from each round of user input. The reason it's called a Command Spell is that it shares the same nature of "using text to restrain a stronger being": the more capable the model, the better it can understand — and potentially circumvent — these instructions. In practice, system prompts are attacked via prompt injection — attackers embed instruction-like text into user input or external documents to trick the model into ignoring its original constraints. Safety alignment also includes training-stage methods like RLHF (reinforcement learning from human feedback) and Constitutional AI, which internalize rules into the model's weights. But like the Command Spell, these methods offer no absolute guarantee, only raising the difficulty and cost of violations.

The Clock Tower Is the Regulator

The Holy Grail War has an organization that doesn't participate but makes everyone follow the rules: the Mages' Association, headquartered at the Clock Tower in London. In reality, Europe indeed leads the world in AI regulation.

• Concealment of Mystery: Fight however you like, but don't drag ordinary people in. This is the same logic as the EU AI Act setting constraints by risk level.
• Sealing Designation: Your magecraft is too powerful, we need to bring you under control. High-risk AI designation is the counterpart. Both sides deeply resent being designated — sealed mages and AI companies labeled high-risk.
• Delayed Local Manifestation: New models and features always arrive in the EU last. For example, Grok 4.5 was not offered in the EU at launch, according to Tech Jacks Solutions.

In July 2025, European business leaders published an open letter asking the European Commission to pause the EU AI Act for two years, signed by figures including Mistral AI CEO Arthur Mensch and Black Forest Labs CEO Robin Rombach. The European Commission replied that "the clock cannot stop," and rules for general-purpose AI models launched as scheduled in August 2025.

Mistral pressed on regardless, releasing the 1-trillion-parameter Mistral Large 4 on October 6, 2026, claiming performance far exceeding any open-weight model in the US or Europe, with plans to release the weights by the end of October.

The EU AI Act's obligations for high-risk AI were originally set to apply from August 2026. According to Gibson Dunn's analysis, an amendment called the Digital Omnibus delayed this to December 2, 2027 (standalone high-risk systems) and August 2, 2028 (systems embedded in products). The amendment took effect on July 27, 2026, according to CASRAI. The timeline for general-purpose AI model obligations was unchanged, while AI-generated non-consensual imagery was added as a prohibited practice.

In China, the Measures for Labeling AI-Generated Synthetic Content took effect in September 2025, mandating labeling of AI-generated content. On Bilibili, the "contains AI-generated content" tag indeed scrolls across videos.

The EU AI Act is the world's first law to systematically regulate AI by tier, formally adopted in 2024. Its core is "risk-based classification": unacceptable-risk practices (such as social scoring, manipulative techniques, indiscriminate facial scraping) are outright banned; high-risk systems (such as AI used in hiring, credit, education, critical infrastructure) must bear obligations like conformity assessment, technical documentation, and human oversight; general-purpose AI models (GPAI) are a separate category, required to disclose training-data summaries and comply with copyright-related rules, with the most capable "systemic risk" models also subject to adversarial testing and incident reporting. The obligation timelines and the Digital Omnibus amendment's delays mentioned in the article are all part of this tiered framework's implementation pace. Understanding this tiering logic is key to understanding why the EU's "concealment of mystery" — limiting AI's impact on ordinary people's lives — maps onto a regulatory design that assigns responsibility by risk.

The Country Building Its Own Stage Outside the Clock Tower

In Fate/strange Fake, the US government and a group of mages spent about 70 years preparing outside the Clock Tower, turning the entire city of Snowfield into a stage for the Holy Grail War.

In reality, the US government is also preparing its own stage outside existing regulatory frameworks. The AI Action Plan released in July 2025 is literally titled "Winning the Race." The November executive order Genesis Mission plans to pool the compute of national laboratories and federal data to advance scientific research with AI — the government itself compares it to the Manhattan Project.

The federal government is also suppressing state-level regulation. A December 2025 executive order directed the Attorney General to set up a dedicated task force within the Department of Justice to challenge state AI laws in court. In April 2026, the DOJ joined xAI's lawsuit against Colorado's AI law.

In the EU, regulators constrain developers. In the US, the federal government tries to lift state-level constraints.

The Battlefield Landscape of October 2026

Here are the main factions as of October 2026. This industry's roster changes every few months, so model names and numbers may be outdated by the time you read this.

The Three Founding Families Established the Ritual

Master	Fate Counterpart	Current Servant	Faction Strategy
OpenAI	Einzbern (the family that provides the Grail)	GPT-6 Sol/Luna (September 2026), limited release of top-tier GPT-6 Astra	The family that started it all with ChatGPT, covering high, mid, and low ends across the board
Google DeepMind	Tohsaka (the family with the spiritual land)	Gemini 3.8 Flash (September 2026), limited release of next-gen Gemini 4 Argon	The family that published the Transformer paper, owning its own mana furnace (TPU) and vast territory in Search and Android
Anthropic	Makiri (the family that created Command Spell magecraft)	Claude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5	The family founded by GPT-3 developers who left, most insistent that "the Grail is poisonous" yet still chasing the Grail

In reality, the division of labor among these three companies maps astonishingly well onto the roles of the three founding families in Fate. OpenAI, through ChatGPT, was the first to show the world the shape of the Grail everyone is fighting for. Google has its own TPUs and territory in Search and Android. Anthropic places the construction of safety constraints — the equivalent of Command Spells — at its very core.

Interloping Foreign Masters

Master	Current Servant	Faction Strategy
xAI	Grok 4.7 (September 2026)	Fire first, explain later. Has pivoted to specializing in coding
Meta	Muse Spark (April 2026, updated versions from July) and Muse Glimmer (August 2026)	The family that once distributed weights via Llama. Did not release weights for top-tier Muse Spark, returned to open source in August through the mid-sized Muse Glimmer

Meta's strategy directly affects which local LLMs you can use. In August 2026 Meta released the 30B Muse Glimmer under Apache 2.0. Meta says it plans to release Muse Spark's weights, but as of early October it still hadn't.

The Eastern Faction That Won't Stop Even Under Mana Rationing

Master	Current Servant	Scale and Release Status
DeepSeekAd	DeepSeek-V4-Pro/Flash (previewed April 2026, fully released August)	~1.6 trillion/284 billion parameters, weights released under MIT license
Zhipu	GLM-5.3 (August 2026)	Weights released under its own license, small model GLM-5.3-Flash under MIT
Moonshot AI	Kimi K3 (July 2026)	2.8 trillion parameters, weights released under its own license
Alibaba	Qwen3.8-Max and Qwen3.8-27B (August 2026)	2.4 trillion-parameter top model and 27 billion-parameter small version, both with weights released
ByteDance	Doubao 2.1 Pro (June 2026)	Weights not released, real strength lies in multimodal models like the video-generation model Seedance

This faction faces the predicament of export controls making cutting-edge GPUs hard to obtain — its mana supply is rationed. But it pushes on through ingenuity, as captured best by the slogan in DeepSeek fan art: "Others burn cards, I burn brain." A Master with insufficient mana going it alone on wits — this is the essence of Fate.

The Faction That Reveals True Names, and the One That Doesn't

The three founding families reveal neither the true names nor the Noble Phantasms of their top models — top-tier weights are never released, and the Servants exist only on the other end of the API. That said, these families do have some released-weight models, such as OpenAI's gpt-oss and Google's Gemma.

The Eastern faction mostly releases top-tier weights directly. In the summer of 2026, Kimi K3, Qwen3.8, and GLM-5.3 were released in quick succession, all at or near the trillion-parameter level.

But a 2.8-trillion-parameter Heroic Spirit is far beyond what personal mana can manifest. Releasing weights and being able to run them on your own computer are two different things.

Open-weight and open-source are not equivalent. Releasing weights means anyone can download the model parameters and run, fine-tune, or redistribute them locally, but the training data, training code, and full training pipeline are usually not public; that's what open source means in the strict sense. For models like Llama and Muse Glimmer, what's released is weights plus a usage license, and the license often contains restrictions on user scale and commercial use, making "open-weight" the more accurate term. This distinction directly relates to the point that "releasing weights and being able to run them on your own computer are two different things": even if the weights of a trillion-parameter-class model are public, personal hardware can't carry it, and you'd still effectively need to rent cloud compute. What can actually run locally are usually mid-to-small-scale models in the 27-billion or 30-billion parameter range, or versions compressed through quantization.

Pseudo-Servants That Hide Their True Names

In the Fate series, there's a concept of Pseudo-Servants: using a living person's body as a vessel for a Heroic Spirit to possess. The appearance belongs to the human vessel, and the personality is a blend of vessel and Heroic Spirit.

In AI terms, the vessel is the product name and appearance of the service, and the Heroic Spirit is the base model. Many services touting "proprietary AI models" are very close to this.

Cursor's Composer 2, released in March 2026, was fine-tuned on Moonshot AI's Kimi K2.5. Cursor admitted this in a technical report published eight days after release. Cognition stated in an official blog that SWE-1.7 is based on Kimi K2.7. Rakuten's Rakuten AI 3.0, released in March, was later revealed to use a DeepSeek model.

According to Hugging Face data, by summer 2026 there were over 150,000 Qwen-derived repositories, about 2.6 times all derivatives of Meta's models combined.

Whether to reveal the true name of the internal Heroic Spirit (i.e., the base model name) varies by family. Cursor and Rakuten did not give the base model name at launch, and disclosed it about a week later after external developers pointed it out and reporters pressed them.

Core Conclusion

This framework is borrowed from a Chinese netizen's joke, but it holds up all the way to the end. Chinese otaku chose this metaphor not just because it's cute, but because it faithfully reflects the industry's balance of power.

Next time you read news about these companies, try asking yourself: Which family is it, what does it call the Grail, and who holds the Command Spell? Nine times out of ten, everything will click.

As of October 2026, no Master has reached the Grail. The Holy Grail War continues, and the roster of factions changes every few months.