Zhigu Trend (ID: zgtrend) | Wang Yanhe
01
$500 billion shock
Storms always come unexpectedly.
On September 11, Tesla's stock price jumped 10%, and its market value increased by 80 billion US dollars, or about 580 billion yuan, overnight.
Traders are horrified at what they see.
The reason for the sudden rise was not Musk's new 370,000-word biography, but a 66-page report released by Morgan Stanley.
Wall Street analysts aren't focused on the new Model 3 or the Mexican Gigafactory, but on the Dojo supercomputer.
Fearing that others wouldn't understand, Morgan Stanley drew a big picture: Dojo is valued at over $500 billion!
What is a dojo? Tesla's AI computing infrastructure for autonomous driving.
Developing autonomous driving requires collecting billions of kilometers of travel data and optimizing train algorithms.
With the advancement of various levels such as L4 and L5, the demand for data and computing power for autonomous driving is increasing exponentially.
To improve training efficiency, Tesla developed its own supercomputer and officially announced it at AI Day 2021, called Dojo. In order not to rely too much on NVIDIA, we even developed our own D1 chip.
Dojo's computing power reaches 1 EFlops, or a billion floating-point operations per second.
Tesla will spend an additional $1 billion to scale to 100 EFlops in October 2024, equivalent to the computing scale of 300,000 NVIDIA A100 chips.
One can't help but think of Amazon Web Services, the world's largest cloud service provider, AWS.
In the beginning, Amazon reserved a large number of redundant servers for its e-commerce and Black Friday shopping festival, and the idle computing power was packaged and sold.
At present, AWS Cloud contributes 70% of Amazon's net profit, far surpassing e-commerce retail and becoming its most profitable business.
Will Dojo be Tesla's AWS?
If Tesla also exports its computing power, self-driving technology, vision algorithms, and AI capabilities, it will gain access to a more valuable ecosystem than cars.
Thus, Morgan Stanley writes:
Dojo gave Tesla an "asymmetric advantage" in its $10 trillion self-driving business.
Obviously, not all automakers are Teslas.
U.S. auto giants such as Ford and General Motors faced a strike of 13,000 auto workers. Workers are demanding higher wages and fear they will be "harvested by new technology." Because the automated production of electric vehicles will put a large number of traditional workers out of work.
Global automakers are also in the throes of slow transformation, price competition, and profit cuts.
That said, few people can build computing infrastructure with hundreds of millions of dollars in cash flow regardless of cost. Huawei estimates that Level 4 autonomous driving will require at least 1 billion kilometers of road testing. That's the equivalent of 1 million cars driving 10 hours a day for a year.
Xiaopeng once revealed that the annual computing power expenditure exceeds 1 billion yuan; The computing power demand of domestic autonomous driving companies has also doubled in the past two or three years.
A consensus is emerging among automakers that it is better to work deeply with cloud service providers than to do it themselves.
Why not band together and become a safer community?
02
Four wheels + supercomputer
Building cars used to be called "four wheels plus a sofa";
Now it's four wheels plus a supercomputer.
As the next generation of mobile terminals, smart cars are all in the cloud, repeating the opportunity of "smart phone + mobile Internet".
Tesla wants to dominate the era of self-driving, and the only opponent he can fight can only be born in China.
The Chinese market leads the world in electrification penetration;
In the first half of 2023, China's automobile exports surpassed Japan, becoming the world's largest automobile exporter.
According to the China Automobile Association, Continental's automobile exports from January to July reached 2.533 million units, a year-on-year increase of 67.9%, a record high. It is expected that exports of new energy vehicles will steadily exceed 1 million units during the year.
Major domestic cloud service providers such as HUAWEI CLOUD, Alibaba Cloud, and Baidu Cloud have also launched automotive cloud services.
With a larger market, a closely coordinated industry chain, and the participation of many developers, China has all the reasons for its birth: a super-large-scale car cloud and cutting-edge autonomous driving technology.
For example, Huawei has established a cloud data center in Ulanqaba and opened a dedicated area for the automotive industry to provide secure, compatible, independent innovation, and powerful cloud computing infrastructure for autonomous driving scenarios.
HUAWEI CLOUD Ascend AI cloud service has been deployed in the automotive field. The kcalorie training will not be interrupted within 30 days, and the breakpoint recovery time will not exceed 10 minutes.
This is the main mode of cooperation between car companies and cloud service providers.
Automotive companies acquire computing power resources in the same way as hydropower dispatch. Cloud service providers address the efficiency of computing power, storage, and network traffic.
However, the cloud needs of automotive companies will also shift from pure computing infrastructure to software services and platform ecosystems.
HUAWEI CLOUD has cooperated with automakers for a long time to build an autonomous driving R&D platform, which can provide two deployment solutions: "one-stop" and "building block".
For start-ups, HUAWEI CLOUD can provide a complete data closed-loop and autonomous driving expert team to help automakers quickly build a complete autonomous driving R&D platform from 0 to 1.
Of course, the development of each car company is different, and the needs of cooperation are of course different.
For example, some car companies want a "sense of boundary" and are unwilling to deeply bundle with cloud vendors; Automotive companies with their own data centers need software collaboration.
In this case, the "building block approach" is more flexible. HUAWEI CLOUD provides an open and open source platform, integrated reference code, and multiple tools that automakers can choose according to their needs. "The key factor influencing the implementation of high-end autonomous driving is no longer to solve common general situations, but to allow models to quickly learn a variety of unusual but constantly emerging long-term problems." Jin Xin, an AI expert at HUAWEI CLOUD, explained.
This is currently the biggest problem facing autonomous driving companies.
Current autonomous driving technology can handle 95% of driving scenarios, but the remaining 5% of edge or turning scenarios are unlearned scenarios.
Collecting tens of thousands of samples to identify new croissants can take weeks or even months.
In theory, at least 10 billion kilometers of road test data would need to be accumulated to achieve fully autonomous driving. The cost and time are unacceptable.
To solve this problem, HUAWEI CLOUD's large-scale Pangu car model can generate virtual space based on data collected by multiple roads.
The object, position, spatial distribution, motion trajectory and other parameters of traffic participants in the virtual space can be adjusted.
The solution can reduce data acquisition and training time in extreme situations from weeks to two days, significantly improving efficiency.
This is also the first time that autonomous driving has used the capabilities of generated data, and the capabilities of Pangu's large model do not stop there.
Cloud service providers are often new to industrial manufacturing.
But Huawei itself is a representative of advanced production. Pangu's large model can cover all scenarios of car companies from automobile design, production to marketing, so that every employee has an AI assistant.
For example, FAW Jiefang is deeply integrated into the Pangu model.
In the past, when designing the look of a car seat, it took them two weeks to complete the rendering. They now use large models that can be generated in seconds and iterate quickly until they are satisfied, realizing that "what you say is what you get."
Another example is more than 80,000 automotive design standards and 1.6 million pages of manuals. It takes weeks for novice designers to sort through the standards, but large models can quickly find relevant chapters and provide source of standards.
After integrating the Pangu model into FAW Jiefang's data, it can provide all-round reinforcement for Jiefang's research, production, supply and sales services.
Huawei and Tesla have in-depth layouts in the field of autonomous driving and AI computing power.
As China's top technology supplier, Huawei has more to bear.
The cart model is a key corner of the board, but the direction of the entire board needs to be seen:
And isn't it just car companies that suffer from computing power anxiety?
There are 40 million companies in thousands of industries. Who doesn't need a smart transformation to regain the certainty of growth?
03
Reopen the world
In 2 months, ChatGPT will turn one year old.
Since July, a total of 130 large-scale models have been launched in China. Who can go next?
From the battle of 100 models to the emergence of bubbles, the new consensus is that the big model focuses on industrial landing.
Studying AI requires algorithms, computing power, and data, but each one is a waste of money.
Not all tracks are like autonomous driving, and automakers and consumers alike will pay for it. Investors are tightening their wallets, defending cash as king and being extra cautious about their moves.
Even OpenAI is showing signs of fatigue, launching an enterprise version of ChatGPT to compete for the B-end market, hoping to make profits faster.
A new world was coming, and financial metrics, technical difficulties and cost control brought him back to reality.
After the "ChatGPT moments" filter faded, businesses also saw the limitations of large models.
For example, the accuracy of large models is not enough. 80% may be enough for everyday use, but not enough for businesses, especially in industries with near-zero fault tolerance, such as legal, pharmaceutical, and finance.
For example, the deployment cost and technical barriers to deployment of large models are too high. Businesses always want low-cost, ready-to-use products and comprehensive services.
Thousands of industries clearly have real needs and real problems;
AI companies also want to land this industry and obtain data flow and cash flow;
But there is a gap between them – between technology and industry.
Huawei has always believed in solving problems, doing complex things, and taking root in the industry.
For example, this year, from the northeast to Beijing-Tianjin-Hebei, Guangdong, Hong Kong and Macao to eastern China, there have been heavy rains throughout the country.
"Heavy rain of the century", "once in a century", "annual rainstorm" and so on are all headlines on social media, affecting millions of people and causing countless economic losses.
Someone asked if our weather forecasting system works? Why can't we give early warning and allow time to prevent disasters?
This is the gap between technology and reality.
Modern weather forecasting technology, simply put: collect a large amount of meteorological data, input it into a supercomputer, and use complex algorithms to simulate forecast results.
Does this seem to be an area where AI excels? That's right.
This summer, the HUAWEI CLOUD team worked around the clock to update the Pangea weather model.
Pangu predicts the path of Typhoon Dusuri
It not only accurately predicted the path of this year's typhoon, but also realized the rainfall forecast for the next 6 hours and 24 hours.
The Pangea meteorological model became the first artificial intelligence model whose accuracy surpassed traditional numerical prediction methods. At present, we have established cooperation with the National Meteorological Administration of China, Shenzhen Meteorological Bureau, European Medium-term Weather Forecasting Center, and Thai Meteorological Bureau.
Just one server is needed to predict the global path of a typhoon in the next 10 days in 10 seconds;
Next, Pangu will challenge the red warning for heavy rain, adjusted from 3 hours in advance to 24 hours in advance.
It's a lifeline race against time, using human technology to fight the chaos and disorder of nature.
On September 20, Huawei CONNECT 2023 was held with the theme of "Accelerating Industrial Intelligence", aiming to work with enterprises in thousands of industries to move towards an intelligent world. The latest developments in the Pangu model were also revealed.
At the model level, the Pangu model forms a three-layer architecture of 5+N+X. The bottom L0 layer is the five basic large models; The L1 layer is N major industrial models such as automobiles, mining, meteorology, pharmaceutical molecules, government affairs, and digital humans; The L2 layer is a specific model for X business scenarios such as supply chain logistics and typhoon roads.
For example, the development of a new drug has been shown to take 10 years and $1 billion. This is the famous "Double Ten Law" in the medical industry.
Through the Pangu drug macromolecule model, the R&D cycle of the main drug can be shortened to 1 month, and the R&D cost can be reduced by 70%.
Professor Liu Bing of the First Affiliated Hospital of Xi'an Jiaotong University developed a super antibacterial drug with the help of the Pangu drug molecular model.
In recent years, Huawei has dispatched a group of doctors, experts, and scientists to mines, factories, live broadcast rooms, and meteorological observatories.
Only by going deep into the front line can you know what all walks of life need. Innovation is not about working behind closed doors and patting yourself on the head, but about moving forward and solving problems.
We also know that people are more concerned about computing power.
After all, NVIDIA A100/H100 has been cut off in China, and even the "castrated" A800/H800 has been sold at a dizzyingly high price, and it is difficult to find.
In the next 10 years, the demand for AI computing power may increase by 500 times. Can domestic AI computing power make up for the gap?
Huawei is fully prepared for this.
They have established cloud data centers in Ulanqab, Gui'an, and Wuhu to support the use of computing power for ultra-large-scale clusters.
In addition, Huawei has conducted comprehensive internal research from hardware to software, from computing power, operator libraries, AI frameworks to AI platforms. It can help customers smoothly migrate from other platforms to the AI cloud service ecosystem of HUAWEI CLOUD.
Simply put, the artificial intelligence computing power generated in HUAWEI CLOUD's home environment eliminates the worry of "bottlenecks" in computing power.
With the assurance of industry-rooted proprietary technologies and ecosystems, HUAWEI CLOUD's large-scale industrial revolution can gain a firm foothold and open up a new dimension.
04
conclusion
Oppenheimer is not a new weapon, but a new world.
Perhaps, in the face of the big model, we should also say: this is not a new technology, but a new world.
This new world in a hurry, whether you call it the fourth industrial revolution or the intelligent revolution, everything will be changed by artificial intelligence, like the beginning of a new world.
The technology we seek must be disruptive innovation, subverting the old formula and bringing new certainty of growth; It must help ordinary people bridge gaps, not deepen barriers.
Companies that can bridge the gap thrive; Enterprises that can benefit thousands of industries are the strong.
The gears of fate have begun to turn.
Only published in the headline number, all other platforms are porters
The picture comes from the Internet, invaded and deleted.
Original is not easy, welcome to forward!
What is your opinion on this? Welcome to leave a comment in the comment area!