While many discuss the booming 'shrimp farming' industry, a technological revolution is quietly taking place in the desolate Gobi Desert, shaping the very lifeblood of our energy supply.

As easily accessible oil and gas resources near the surface become depleted, modern 'oil hunters'—who once relied on footsteps to measure the earth—are now setting their sights on depths exceeding 10,000 meters. They seek oil layers mere meters thick amid kilometer-deep strata. Traditional experience-based 'blind box' exploration methods have proven ineffective, replaced instead by an invisible 'computing battle'.

To uncover the secrets of this cutting-edge technology, we spoke with Gong Renbin, former chief expert at the PetroChina Exploration and Development Research Institute. Through the lens of a 40-year 'oil IT veteran,' we explore how computing is reshaping humanity's quest for 'black gold'.

After graduating from university in 1986, Gong Renbin dedicated himself to building petroleum IT systems, witnessing China's oil and gas industry leap from informatization to digitization and then to intelligence.

'Old-generation oil explorers relied on sturdy legs and accumulated experience,' Gong recalled, painting a vivid picture of traditional petroleum exploration. Geologists traversed mountains and rivers, constructing subsurface oil and gas structures in their minds through limited surface sampling, visual observation, and lifelong expertise.

This model, heavily reliant on human brains and physical effort, is nearing its cognitive limits. Modern exploration faces two unavoidable challenges.

The first challenge is the exponential—even 'nuclear-level'—growth in data volume.

'When I first started, a surface explosion with a 1,000-meter reception depth, maybe 1,500 meters, was standard,' Gong recalled. Exploration then resembled feeling stones in shallow water. Today, depths reach an astonishing 15,000 to 20,000 meters.

Not only has exploration depth doubled, but data collection dimensions have expanded geometrically. Past explosions deployed 120 ground sensors; now, tens of thousands simultaneously capture faint echo signals, yielding exponential data growth.

'A data volume might have been several hundred megabytes before, then hundreds of gigabytes. Now, even ordinary seismic exploration areas generate hundreds of terabytes, some reaching dozens of petabytes,' Gong said, sharing an 'extreme example': 'Copying 150 TB of raw individual data (unitary data) took nearly a month physically.'

The second challenge is the 'pressure' of massive, complex data on high-performance computing (HPC).

Oil and gas exploration resembles Earth CT scans. More data and finer processing yield accurate subsurface geological insights, increasing oil discovery odds. Without denoising, deconvolution, and stacking, petabyte-scale data holds little value.

Seismic data processing involves dozens of steps, hundreds of software modules, and compute-intensive parallel tasks, demanding precision and timeliness far beyond ordinary computers. Since China's 'Yinhe II' supercomputer, HPC has underpinned geophysical data processing.

By 2026, oil and gas exploration and reservoir simulation rank among Kunpeng HPC's largest application and innovation scenarios. To meet multi-task, massive parallel computing demands, Kunpeng-based HPC platforms with fully distributed storage drastically reduce I/O wait times, satisfying stringent oil and gas industry requirements for speed and parallel computing.

This creates a 'counterintuitive' phenomenon: while intelligent computing and large models dominate discourse, in AI for Science, intelligent computing accounts for just 30% of applications, with most tasks relying on HPC's precise computing.

In oil and gas exploration, only after HPC thoroughly cleanses data can AI large models deeply intervene and identify patterns.

With HPC-generated high-quality data, large models find fertile ground. Yet AI's oil and gas industry trials face unexpected hurdles.

Gong noted that 'artificial intelligence' in seismic data processing appeared in Chinese journals as early as 1984. For decades, 'oil hunters' relied on 'sturdy legs.' An expert needed 1-2 days to interpret a new well, and an oil field drills thousands annually—an immense workload.

The reason is clear. Traditional models focused excessively on algorithms, neglecting data volume.

'I once saw a paper using data from 34 wells to build a model for logging interpretation. It's like a barefoot doctor treating 34 villagers then practicing medicine globally,' Gong quipped, highlighting poor generalization in traditional models.

The turning point came in late 2021.

During a visit to Huawei's Hangzhou Research Institute, Gong was inspired by the Pangu weather model's success: 'If massive meteorological data can train a model significantly improving forecasts, why not use seismic data for an 'earthquake interpretation large model' in oil and gas exploration?'

With clarity, Gong led the PetroChina Exploration and Development Research Institute team, partnering with Huawei, China Mobile, and others, to 'refine' industrial-grade large models.

To ensure 'pure nourishment' for the large model, the team collected approximately 200 TB of real business data across 60,000 square kilometers in seven basins, including Ordos, Tarim, and Sichuan. During initial data organization, HPC and manual efforts combined for 'better safe than sorry' cleansing—removing any inappropriate artificial geological interpretations to prevent 'misleading' the large model.

After 93 days of continuous training, an 8 billion-parameter vertical large model emerged, swiftly deployed in frontline operations.

In Nanchong, Sichuan, facing a 25,600-square-kilometer mega-dataset, traditional cross-term identification proved time-consuming. The large model completed the task in 10 days, achieving approximately 90% accuracy and 9x efficiency.

For unconventional tight gas forecasting in Changqing Oilfield, without well data, it delivered results in a week versus three months traditionally.

Interpretation work that once took six months now takes two weeks at most, boosting efficiency 9-12x...

What impressed Gong most was the large model's 'counterintuitive' oil-finding ability.

Traditional exploration relied heavily on experts' subjective experience—once invaluable, now potentially creating blind spots. The large model, a 'pure materialist,' excels at 'mining' vast data, capturing oil reservoir signals missed by expert experience.

The shift from 'experience-led' to 'data-driven' marks a new era: nurtured by HPC and high-quality data, AI's seeds have 'taken root' kilometers underground, ushering in an age of 'cloud-based oil calculation' from 'field-based oil hunting'.

Computing power can be bought, models jointly developed, but the toughest intelligentization (intelligent transformation) hurdle is bridging the knowledge gap: business experts lack AI knowledge, and AI specialists lack oilfield expertise.

Gong identified the energy industry's most genuine, yet frustrating, pain point: 'Honestly, our oil and gas salaries can't attract top AI talent,' proposing a pragmatic 'three-tier talent pyramid' strategy.

Tier 1: Business-savvy 'Lampbearers'.

Select senior business personnel interested in AI internally. They may not code but deeply understand oil and gas exploration, accurately defining problems and core needs.

Tier 2: Internally trained 'Tuners'.

If AI experts are unattainable, identify math, physics, or geophysical-savvy young staff for systematic AI training, debugging open-source algorithms, fine-tuning conventional models, and gradually becoming AI implementation pillars.

Tier 3: Externally collaborative 'Joint Teams'.

For foundational innovation and architecture R&D, partner deeply with Huawei, Alibaba, iFLYTEK, and universities like Tsinghua and Peking, merging external general AI with internal industry expertise.

Beyond systemic talent solutions, Gong warned against industry chaos in blindly following large models.

As a veteran of multiple national oil informatization (informatization) projects, Gong outlined four common 'pitfalls' in intelligent transformation: pursuing 'smart' for its own sake, detached from production; blindly adopting large models; prioritizing technology over business; and severe duplication.

'Blindly pursuing large models' is a key concern, Gong emphasized: 'Without clear needs or high-quality data, jumping into large models without purpose leads nowhere.' Weak data foundations only 'mislead' models.

Gong's views align with Chinese Academy of Engineering member Zheng Weimin's.

At an industry conference months ago, Zheng proposed 'HPC+AI for Science,' using petroleum exploration to illustrate the new research paradigm from HPC-AI synergy: HPC handles data preprocessing, while large models predict oil presence.

Behind this rationality lies a new trend in digital transformation across industries: pure hype recedes, replaced by deep integration.

Examples include HPC-AI base fusion computing—Kunpeng HPC cleans data front-end, large models predict intelligently back-end, forming a perfect computing loop—and industry-academia-research collaborative talent cultivation, breaking barriers between theory and practice to nurture AI-industry hybrids...

There's reason to believe that when business 'Lampbearers' illuminate the path, 'Tuners' and 'Joint Teams' align technology, and 'HPC+AI' computing bases solidify infrastructure, the AI-industry divide will be 'flattened'.

The underground 'computing battle' cools today's AI frenzy, charting a clearer path for industries.

Large models are neither 'plug-and-play' nor cure-alls. Unleashing their 'magic' in frontline operations requires HPC to cleanse petabyte-scale chaotic data, business experts to clarify needs, and organizational systems to support 'data-to-decision' engineering.

As Gong put it: 'Success requires two-way effort, not unilateral tech push. Not bigger models, but deeper integration.'