Folks, you know that nagging feeling I'm talking about —

Following the AI news in the first half of this year has been like binge-watching a TV series.

In January, Company A claims its model is the best. February rolls around, Company B casually overtakes them. By March, Company C comes out of nowhere and steals the crown. The top spot on every leaderboard changes hands before the quarter's even over.

And the craziest thing? Every time the crown changes hands, a chorus chimes in: "Scaling Law has hit a wall." "Large language models are done for."

Then a new benchmark drops, and suddenly that chorus goes quiet.

Until next month's winner shows up, and the exact same doubts come crawling back.

Honestly, I've been following this for six months, and I'm exhausted.

But when you step back and look at the full picture, one thing becomes crystal clear —

This isn't "AI hitting a dead end." On the contrary, the first half is over — and the second half has only just begun.

And in the second half, the rules have completely changed.

Scaling Law Isn't Dead — It Just Stopped Playing the "Stack Parameters" Game

Let's start with the thing that's been questioned a thousand times over: Scaling Law.

In plain English — bigger models perform better. Throw more parameters, more data, more compute at the problem, and capability scales up. That's it.

OpenAI proposed this back in 2020, and for years everyone's been burning cash following that formula.

But this year hit a snag. Anthropic's Opus series started showing diminishing returns. GPT-5.1 through 5.4 showed only marginal improvements. So the "Scaling Law has hit a wall" crowd crawled out of the woodwork.

Let's be real — that's like saying "the only way to build muscle is squats, and since I've hit a squat plateau, working out is useless."

Sure, single-dimensional parameter stacking has diminishing returns. But the AI labs have already pivoted to three completely new directions to scale —

① Multimodality

Used to be, LLMs just read text. Now they need to see images, hear audio, understand video — all at once. Google's Gemini 3 briefly took the crown precisely because it baked text, images, audio, and video into the same vector space during training — not stitched them together as an afterthought.

DeepMind went even further. It built something called Omni and straight-up called it a "world model" — one that can understand physical rules AND generate causally consistent visuals.

Translation: AI used to grow up in a library. From now on, it's growing up in the real world.

The implications are absurdly vast. Gravity and collision patterns learned from video can feed back into optimizing physics simulation code; the recursive logic from code can make video generation more coherent. Different modalities infect each other.

One fascinating detail: the global text-to-video leaderboard's top ten now includes nine Chinese models. The top spot belongs to ByteDance's Seedance 2.0, followed by Alibaba's Wan at number two and Kuaishou's Kling 3.0 at number three.

But let's be honest — China leads on the "generation" side but is still playing catch-up on the "understanding" side. Making great videos doesn't mean you can comprehend what's happening in them.

② Multi-Agent Systems

Since making a single model smarter is getting slower, just have multiple agents work together.

Moonshot AI's Kimi Work is the most extreme example — 300 agents working in parallel, running for 13 continuous hours, autonomously calling 4,000+ tools across the full pipeline: data cleaning, code writing, testing, documentation. The whole thing.

Humans work sequentially. AI can work in parallel. That's a dimensionality reduction attack on traditional work itself.

Of course, there are headaches. 300 agents running simultaneously burn tokens at a painful rate. One agent failing can collapse the whole task chain. And when a bug shows up, good luck figuring out which link in the chain caused it.

③ RSI — Recursive Self-Improvement

This name sounds intimidating, but it's really just letting AI train itself.

Anthropic dropped a data set in June that I had to read twice to believe —

80%+ of Claude's codebase is now written by AI itself. Engineers' daily merged code output is 8× what it was last year. A single AI agent independently completed an AI safety research project — 800 hours of cumulative work — outperforming a week's worth of human researcher output.

AI writes code to train better AI, and that better AI writes even better code.

This flywheel is already spinning.

To be clear, letting AI fully determine its own model architecture, parameter scale, and training strategy is still a ways off. But the direction is unmistakable.

The Throne Passes Around — Nobody Can Keep It Warm

With the fundamentals covered, let's look at who actually won the first half.

I'll keep it simple — a timeline is all you need:

?️ Phase 1: Google Leads (Late 2025 – Early 2026)

Google Gemini 3 briefly led the pack, powered by native multimodality, YouTube's massive video corpus, and Waymo's physical interaction data. Data moats were its trump card — advantages OpenAI and Anthropic couldn't replicate quickly.

? Phase 2: Anthropic Surges (Q1 2026)

Claude Code exploded in the developer community, and the product flywheel started spinning:

Great product → attracts developers → generates real coding data → feeds reinforcement learning → stronger model → better product.

By May, Anthropic's annualized recurring revenue hit $47 billion, surpassing OpenAI in both absolute value and growth rate.

? Phase 3: Anthropic Stalls, OpenAI Strikes Back (Q2 2026)

But Anthropic's honeymoon lasted exactly one quarter.

Compute capacity hit a wall. Opus 4.7 iterations slowed, the Fable model was launched then pulled, and they even started raising prices and throttling usage.

OpenAI seized the window and dropped GPT-5.5 and 5.6, matching or surpassing Claude on coding benchmarks. What was once criticized as "excessive compute reserves and runaway CapEx" became their biggest weapon in the price war.

Throw in Musk's Grok 4.5 and China's DeepSeek V4 as constant disruptors, and the gap between the top-tier players has shrunk to a matter of months.

The Scorecard

Whoever stabilizes in the second half won't win on parameter count alone — that game ended long ago. It's about who can simultaneously nail model capability, product experience, data flywheel, and compute supply.

Open-Source Is Eating Closed-Source's Lunch — And Getting Greedier By the Day

The global AI race had a clear stratification: US premium closed-source (OpenAI, Anthropic, Google) charging API fees for service revenue, and Chinese open-source (Qwen, DeepSeek, Zhipu, Kimi) using permissive licenses to aggressively court developers worldwide.

This year, that stratification shattered.

The reason is simple — the stronger and more expensive Anthropic and OpenAI's flagship models got, the more enterprise customers looked for alternatives.

The result: enterprises have developed a brutally pragmatic tiering strategy:

A thing called a "model router" has become standard infrastructure: enterprises submit a task, the system automatically judges complexity, routes simple work to cheap DeepSeek and Qwen, and only calls expensive closed-source models for complex tasks.

The numbers tell the story best.

On OpenRouter, DeepSeek's token share exploded from 9.1% to 18.1% — doubled, leaping from the fourth tier to global number two. Meanwhile, Google's share cratered from 24.6% to 10.7% — nearly halved.

A year ago, global enterprises were wary of Chinese open-source models — security vulnerabilities, data compliance concerns. Now, many financial institutions are already running them. No security incidents. Just serious cost savings.

The kicker? Even Microsoft is reportedly evaluating DeepSeek for powering Copilot Cowork. Nobody — not even the giants — wants to be held hostage by a handful of closed-source vendors' pricing power.

Chinese AI models' overseas push has crossed the threshold from "usable" to "competitive."

• DeepSeek — cost advantage is ridiculous, the most mainstream Chinese model in overseas open-source markets

• Zhipu GLM-5.x — coding capability so strong, overseas vertical apps like Harvey train dedicated models on top of it

• MiniMax M3 — breakthrough on native multimodality

As global enterprises pivot from "token maxxing" to cost optimization, Chinese models' cost-performance advantage will only grow.

Model Makers Building Apps, App Makers Building Models — Lines Are Blurring

Something fascinating happened this year: upstream and downstream have started invading each other's turf.

? Downstream Push: Anthropic vs Figma

On one side, model companies like Anthropic are pushing aggressively downstream. Claude Code goes head-to-head with Cursor and Codex in the coding arena. In April, they dropped Claude Design — an AI design tool that went straight for Figma's jugular. Figma's stock dropped 7% on launch day.

The shift from selling API tokens to selling industry solutions is a business model upgrade — call-volume revenue has a hard ceiling, but value-based pricing for business outcomes opens up vastly more space.

? Upstream Push: SpaceX + Cursor

On the other side, application companies are pushing upstream.

The most explosive case: In late April, SpaceX acquired Cursor for $60 billion in an all-stock deal. By June, Cursor announced plans to build its own 1.5-trillion-parameter foundational model, trained on SpaceX's supercomputing cluster — capable of marshaling over 100,000 GPUs.

Translation: Cursor was surviving on Anthropic and OpenAI's APIs — the same two companies that are simultaneously building competing products (Claude Code and Codex). A textbook "your supplier is also your competitor" death trap.

SpaceX's acquisition solved two problems at once:

1. Compute autonomy — no longer beholden to API suppliers

1. Data goldmine — Cursor's real developer interaction data (how code is written, which suggestions are accepted or rejected, debugging patterns) far exceeds the quality of GitHub's static open-source repos

This chess game gets more interesting with every move.

The Three Vibe Waves: Code, Work, Create

LLM commercialization hasn't penetrated all industries evenly. It comes in waves ranked by "verifiability from high to low."

? Wave One: Vibe Coding

Why did coding monetize first? Because code either compiles or it doesn't — yes or no, no gray zones. Reinforcement learning can iterate efficiently, and users can see value immediately.

But H1 2026 saw coding hit a ceiling. When GitHub Copilot switched to per-token billing, enterprise bills exploded:

Uber burned through its entire annual AI budget in four months.Amazon straight-up killed its internal AI tool usage leaderboard.

Enterprises have pivoted from "maximize token consumption" to "control costs."

? Wave Two: Vibe Working

This wave is rapidly penetrating. OpenAI disclosed in June that knowledge workers — spreadsheet jockeys, PPT makers, researchers — now make up 20% of Codex users, growing 3× faster than programmers. Individual user growth is outpacing programmers by 4×.

Finance professionals are Anthropic's second-largest revenue source. Investment researchers use AI for valuation modeling and portfolio review. Marketers use it to read proposals, produce analysis reports, and generate PPTs. Analysts use it to consolidate roadshow notes. The shift from individual subscriptions to corporate reimbursement is clear.

But this wave faces bigger friction than coding:

Problem: A report's quality or a suggestion's validity has no objective yardstick — making RL training harder and user trust thresholds higher. Enterprise contracts, financial data, and HR information handled by AI requires compliance rules — audit trails, accountability, traceability — that haven't been built yet.

Big enterprise penetration will be slower.

? Wave Three: Vibe Creating

ByteDance introduced this concept in June, based on Seedance 2.0's performance leap. The core idea: "tell a good story, return to intent expression." You don't need to master complex cinematography terminology — describe the feeling, and AI generates the visuals.

Chinese models have already reached the global first tier in video generation. Kling can produce cinema-grade performances with layered acting and complex lighting. Seedance skews more toward short-video aesthetics but iterates at breakneck speed.

But commercialization is still orders of magnitude behind Coding and Working. Two core pain points:

1. Gacha rate — the same prompt might produce one usable output in ten tries, with no way to fine-tune precisely

1. Limited scope — use cases are stuck on short-form clips and ads; full-length content production isn't viable yet

That said, Creating's long-term ceiling is the highest of all three.

Coding lets people who can't code build software.Working lets people who can't use Excel analyze data.Creating lets people who never learned video editing or music composition express their ideas.

It doesn't replace creators — it expands the definition of who a creator is.

Three Lines to Watch in H2

? Line One: OpenAI and Anthropic Go Public

Likely sometime between H2 2026 and 2027. Once these two hit the market, the entire AI model sector gets a pricing anchor.

? Line Two: Google Still Has a Comeback Card

Google is behind in the coding track short-term, but it holds two trump cards:

1. TPU cluster + Google Cloud — inference cost advantage will be formidable once fully deployed

1. Multimodality + world models — YouTube/Waymo data + Omni architecture are structural advantages nobody can copy quickly

? Line Three: Chinese Open-Source's Global Share Keeps Climbing

This is the most relevant line for China's domestic industry.

• DeepSeek — algorithmic cost advantage

• Zhipu — coding capability

• MiniMax — native multimodality breakthroughs

All of this is hard currency in the global enterprise cost-optimization trend. The more overseas vertical applications train on Chinese open-source foundations, the thicker the ecosystem moat gets.

In Closing

Folks, the large model industry in mid-2026 is nowhere near the finish line.

Everyone used to think pushing Scaling Law to its extreme would produce a single winner.

Looking back now, parameters are just the entry ticket. Product, data, compute, ecosystem, vertical scenarios — every single dimension is a new battlefield. Leads are shorter-lived than ever. And for those falling behind, comeback opportunities are more abundant than ever.

For China's AI industry, the first half was about playing catch-up. But in the second half — on open-source, multimodality, and the global stage — we're no longer spectators.

This show isn't over.

Who wins? Who loses?

Stay tuned.

3DSTOR is a global IT component supplier that has been established for many years and has established long-term and stable cooperative relationships with top well-known brands such as Intel, AMD, Nvidia, Western Digital, and MSI. The main products include servers, motherboards, graphics cards, CPUs, hard drives, etc.We have long been focused on serving the global B2B market, providing a variety of AI/ML/HPC solutions, and can provide one-stop procurement services for IT hardware customers with different needs. We promise that our products are 100% brand new and original, and we will also conduct technical testing on the products before shipping.