The "Trust Crisis" of Large Models Erupts in the Second Half of 2026: We're Paying Twice – Once to Let AI Write, Once to Prevent AI Fraud

The summer of 2026 is witnessing an unprecedented "aesthetic fatigue" and "trust crisis" in the global large language model industry. Just over the past week, the tech world was jolted by a late‑night bombshell – OpenAI's Chief Scientist gave an internal sneak preview of the upcoming GPT‑5.6, which is set for an emergency launch at the end of June. At the same time, email giant Superhuman announced its acquisition of GPTZero, the world's largest AI text detection unicorn, at a premium, aiming to build a "authenticity defense line" within its massive enterprise email traffic specifically to intercept AI‑generated content. A bizarre yet real industry inflection point has taken shape in the second half of 2026: on one hand, we are paying hefty subscription fees to have large models generate code, documents, and emails at scale; on the other hand, we have to spend even more expensive tokens to hire another AI to detect and clean up the "hallucinatory garbage" produced by the first AI. This surreal "AI‑vs‑AI" meat‑grinder battle is ruthlessly piercing the last veneer of warmth around the large model industry.

I. GPT‑5.6's "Defensive Release": The Fading of Silicon Valley's Technological Myth

To understand this meat‑grinder battle, one must first see the real anxiety over the cloud computing ceiling. In the first half of 2026, Silicon Valley's narrative around AGI (Artificial General Intelligence) began to show signs of diminishing returns. The depletion of data dividends, combined with soaring computational power and electricity costs, has made the underlying iteration of large models increasingly difficult. Training a trillion‑parameter large model now costs three times more in electricity than it did two years ago, while high‑quality text data on the internet is nearly exhausted, and synthetic data carries the risk of self‑contamination. These hard constraints have forced Silicon Valley giants to rethink the old "brute force" approach.

OpenAI's decision to preview GPT‑5.6 in secret is, in essence, a "defensive release." In the latest industrial‑grade coding ability tests this year, the much‑anticipated GPT‑5.5 failed to create a generational gap and was frequently matched by both Anthropic and top open‑source models from China. Once Silicon Valley's technological myth fades, a capital market re‑evaluation inevitably follows. Investors are now asking: "Where is the next killer application? Will linear growth in model parameters still yield exponential returns on commercial investment?" Under this pressure, OpenAI had to deliver a reassuring answer to the market before the end of June.

On June 27, Beijing time, OpenAI officially released the GPT‑5.6 family of three models: Sol (Sun), Terra (Earth), and Luna (Moon). The flagship model, Sol, is positioned for the most complex reasoning, scientific research, software development, cybersecurity, biological research, and AI agent workflows. In the Terminal‑Bench 2.1 benchmark for programming workflows, Sol scored 88.8% in standard mode and reached 91.9% with Ultra mode enabled. While this performance leads its predecessor, the advantage over Anthropic's Claude Opus 4.6 has narrowed to less than two percentage points. The context window of GPT‑5.6 has expanded from 1 million tokens to 1.5 million tokens – a 43% increase – enabling it to process entire codebases at once, a clear boon for large‑scale enterprise code maintenance.

However, even more striking than the model's performance is the release method – OpenAI disclosed in its announcement that, at the request of the U.S. government, the model would first be made available in a limited preview to a small group of "trusted partners," and some customers' access to the model even requires individual approval by the U.S. government. The shift from "voluntary frameworks" to "one‑customer‑one‑review" signals that the global deployment of large models is facing unprecedented geopolitical barriers. European regulators are also advancing the AI Liability Act, requiring high‑risk AI systems to pass stringent interpretability tests. This means that no matter how capable a model is, if it fails to meet compliance requirements, it will struggle to enter mainstream commercial markets. The tension between technological leadership and regulatory compliance is reshaping the entire industry's pace of development.

II. "AI Detecting AI": The Most Absurd Business Loop of 2026

In the second half of the large model era, the biggest headache for business leaders is no longer "AI isn't smart enough," but rather "AI is creating epic‑scale information garbage and hallucination pollution." The financial sector is the first to suffer – an AI‑generated research report containing fabricated financial data could lead to millions of dollars in misjudgments; the medical field is equally vulnerable – if AI‑assisted diagnostic suggestions include hallucinated drug names, patient safety is directly threatened. For this reason, enterprise trust in AI‑generated content is rapidly crumbling.

Superhuman's premium acquisition of GPTZero completely tears away the veil of the AI efficiency myth. Originally a tool for educational institutions to check whether student essays were AI‑written, GPTZero has quickly pivoted to become an enterprise‑grade content risk‑control platform. Its technical principle is not overly complex: it determines whether a text has machine‑generated statistical characteristics by analyzing its "perplexity" and "burstiness." But the difficulty lies in the fact that as large models write ever more like humans, both false‑positive and false‑negative rates are climbing. To improve accuracy, GPTZero has to continuously iterate its own detection model – even using more powerful large models to train its discriminators – creating an inward‑spiraling "AI‑against‑AI" arms race.

As industry observers have noted: "When you open your corporate email every morning and find 100 impeccably polished reports generated by employees using Claude, plus business proposals auto‑generated by ChatGPT from suppliers, productivity doesn't rise – it completely paralyses." Because no one can guarantee that these highly professional texts do not contain hallucinated facts fabricated by the large model. Internal tests at a top investment bank showed that among unverified AI‑generated reports, about 3.7% of statements contained factual errors, and over half of those errors could not be quickly identified by the naked eye. This has given rise to the most absurd and most profitable business loop of 2026:

• The first payment: companies buy large‑model accounts for employees and deploy agents, continuously generating automated content.

• The second payment: companies then purchase AI detection tools like GPTZero to scan, filter, and cleanse AI texts submitted by their own employees, much like antivirus software.

Large models have increased content production speed by ten thousand times at extremely low cost, yet human society is forced to place its most expensive trust currency back into the hands of "AI judges." According to the latest statistics from market research firms, the global market for AI content detection and governance software reached $4.7 billion in Q2 2026, with a year‑on‑year growth rate of 210% – far outstripping the growth rate of the large model market itself. This may be the most biting satire on "technological neutrality" – the tools we invent ultimately require another tool to supervise them.

III. Chinese Large Models Break Through Against the Odds: Doubao 2.1 Crosses the "Quality Inflection Point"

While Silicon Valley is mired in a "trust crisis," Chinese large models are accelerating their breakout. On June 23, Volcano Engine held its 2026 Summer FORCE Original Power Conference in Beijing, releasing Doubao large model 2.1 Pro, along with several new models for video, image, and audio. Doubao 2.1 Pro achieved significant capability leaps in Coding, Agent, and VLM (Vision‑Language Model) directions, with multiple evaluations showing performance superior to Claude Opus 4.6 – officially crossing the production‑grade "quality inflection point." This inflection point means that the model's stability and accuracy in complex engineering tasks has reached a level that can replace junior engineers, rather than remaining merely a "chat toy."

In a chip design test, Doubao 2.1 Pro ran continuously for nearly 18 hours, iterating through 9 rounds, and successfully completed the entire engineering workflow encompassing simulation, testing, and comprehensive verification. It can also simultaneously coordinate over 500 virtual characters, each performing its own role in a 3D virtual city scene, generating over a hundred buildings. Even more noteworthy is its overall usage cost – nearly 80% cheaper than leading foreign models. This cost advantage was directly reflected in market feedback: on the day of release, more than 2,000 enterprise clients applied for internal testing, spanning multiple verticals including smart manufacturing, game development, and financial risk control.

As of June this year, Doubao's daily average token call volume has exceeded 180 trillion, growing more than tenfold over the past year. IDC data shows that in China's public cloud MaaS service market, Volcano Engine holds a 49.5% market share, ranking first. At the same time, Zhipu open‑sourced GLM‑5.2, pushing million‑token context from "paper specs" to "production readiness," handling over 880,000 tokens in a single continuous task. Global large‑model call volume has grown for eight consecutive weeks, while China's large‑model call volume has surpassed that of the United States for seven consecutive weeks – token call volume has become a high‑frequency barometer for AI application penetration. The continued rise of this metric indicates that China's AI application ecosystem is moving from "trial" to "essential," extending from C‑entertainment to B‑core business processes.

The counter‑trend rise of Chinese large models has also spurred coordinated development across the upstream and downstream industrial chains. Model vendors, cloud service providers, and industry ISVs (independent software vendors) are forming tight ecological alliances, jointly promoting AI adoption in scenarios such as industrial quality inspection, supply chain optimization, and smart cities. Unlike Silicon Valley's single‑minded pursuit of parameter scale, domestic players place greater emphasis on "cost‑performance" and "deployability." This pragmatic approach may well help China's AI industry carve out a differentiated rhythm in global competition.

IV. Computing Power Infrastructure: The "Fuel" and Bottleneck of Large Model Surge

Behind the frenetic iteration of large models and the explosive growth in token call volume lies an indispensable pillar – computing power. NVIDIA's revenue grew 65% over the past year, operating profit rose 60%, and free cash flow exceeded $96 billion in fiscal 2026. NVIDIA CEO Jensen Huang continued to reinforce the narratives of AI infrastructure, AI factories, and the "Token economy" at the shareholder meeting. He likened AI factories to a new generation of power plants, with computing power as the electricity – without sufficient computing supply, neither training nor inference of any large model is possible.

At the same time, OpenAI teamed up with Broadcom to release its first self‑developed AI inference chip, Jalapeño, in an effort to more tightly bind models, products, and underlying hardware. Qualcomm also unveiled its data center strategy with the Dragonfly data center product portfolio, securing clients like Meta and Microsoft. Behind this intensive deployment by chip giants in AI inference chips is a rapidly expanding market. According to data from the China Academy of Information and Communications Technology (CAICT), the domestic computing power leasing market reached RMB 68 billion in Q1 2026, and is expected to surpass RMB 260 billion for the full year. This rapid market expansion is rooted in a huge supply‑demand gap – in Q1 this year, domestic AI computing demand grew 417% year‑on‑year, while supply grew only 128%. AI cloud providers are seeing backlog growth, GPU leasing prices are rising, and the booming outlook for computing power, storage, and leasing services is likely to continue.

For the vast number of SMEs and developers, GPU procurement costs that easily run into the tens of millions, coupled with long delivery cycles, make GPU leasing and computing power leasing the optimal solution to lower the barrier to AI adoption. Whether it's pre‑training, fine‑tuning, or inference deployment of large models, flexible computing power leasing models are helping enterprises and developers join this AI race at lower cost and greater speed. At present, domestic intelligent computing centre construction has also entered an acceleration phase – more than 30 cities are planning or expanding their intelligent computing infrastructure. However, the actual operational utilisation rates of these centres are uneven. Some centres suffer from idle capacity due to insufficient scheduling capabilities and incomplete software stacks, while high‑end GPUs in first‑tier cities remain "hard to come by." This structural contradiction highlights the value of professional computing power leasing platforms – they can efficiently match fragmented computing resources to users with urgent needs through elastic scheduling and pay‑as‑you‑go billing, thereby improving overall societal computing utilisation efficiency.

Furthermore, the computing power leasing market also benefits from the surge in inference demand. Training a large model may take only a few months, but after launch, inference services represent a continuous, long‑term demand. With the proliferation of Agent applications and real‑time interactive scenarios, inference computing consumption has already surpassed training computing and become the mainstay of computing consumption. This means that computing power leasing is no longer just a "training‑phase" temporary need, but a recurring expenditure throughout the entire lifecycle of AI applications. Those platforms that can provide stable, cost‑effective GPU leasing services are poised to reap sustained commercial returns in the coming years.

V. Where Is the Industry Headed?

The AI industry highlights from the past week outline several clear directions for the second half of 2026:

First, large models are shifting from a "capability contest" to a "trust contest." As model capabilities begin to overflow, how to identify, filter, and manage AI‑generated content is becoming a new multi‑billion‑dollar arena. AI content regulation is also being fully implemented – the four‑ministry "Measures for the Identification of AI‑Generated Synthetic Content" has entered full enforcement, and major short‑video platforms are mandating AI‑content labelling. In the future, every piece of AI‑generated text, image, and video will need to carry a "digital watermark" for traceability and verification. While this increases compliance costs for content producers, it also provides an institutional foundation for society to rebuild trust in AI.

Second, Chinese large models are accelerating their transition from "usable" to "good‑to‑use." Doubao 2.1 crossing the quality inflection point and Zhipu GLM‑5.2 achieving adaptation with multiple domestic computing platforms on its first day of launch – domestic large models are moving from general Q&A toward developer tools and enterprise workflows. In the second half of 2026, we are likely to see more industry‑specific large models go live – such as financial LLMs, medical LLMs, industrial LLMs, etc. – which will build on common base models and integrate industry‑specific data to create differentiated competitiveness. This will also further drive demand for differentiated computing solutions, such as industry‑optimised GPU leasing packages, which will become a new selling point for computing power leasing platforms.

Third, computing power infrastructure is becoming the "game‑changer" that determines the direction of the AI industry. From NVIDIA's AI factory narrative, to Qualcomm's nearly $4 billion acquisition of Modular to challenge the CUDA ecosystem, to the shift of domestic computing power adaptation from "strategic reserve" to "production necessity" – computing power is moving from behind the scenes to centre stage. We expect that in the second half of the year, more AI chip startups and computing power scheduling platforms will secure funding, and the business model of computing power leasing will evolve from simple "resource rental" to "one‑stop solutions integrating computing, algorithms, and data." Whoever can provide more convenient and cheaper computing services will occupy a favourable position in the coming wave of AI application proliferation.

Large models have increased content production speed by ten thousand times at extremely low cost, and computing power leasing is providing a steady stream of fuel for this transformation in a flexible and elastic manner. As "AI‑detecting‑AI" becomes the norm in the second half of 2026, one thing remains certain – computing power will never be in surplus. Whether it's cloud training, edge inference, model fine‑tuning, or data cleansing, every link depends on underlying computing resources. In the context of long‑term computing scarcity, professional GPU leasing and computing power leasing platforms will become the indispensable "water, electricity, and coal" of the AI industrial chain, safeguarding the intelligent transformation of thousands of industries.