A gray rectangle Rows of flickering machines line buildings on the outskirts of San Jose. Colorful wires connect high-end servers, networking equipment and data storage systems. Bulky air-conditioning units roared overhead. The noise forces tourists to shout loudly.
The building belongs to Equinix, a company that leases data center space. The equipment inside belongs to companies ranging from corporate giants to start-ups, which are increasingly using it to run their artificial intelligence (artificial intelligence) system.this artificial intelligence The gold rush, fueled by the incredible complexity of “generated” systems like chatcommon technology, a popular virtual interlocutor, promises lucrative profits for those who tap into the potential of the technology. Still, like the early days of any gold rush, it has created fortunes for sellers of the necessary picks and shovels.
Nvidia Designs Semiconductor of Choice for Many, May 24 artificial intelligence servers, topping analysts’ revenue and profit forecasts for the three months ended April. It forecast sales of $11 billion for the current quarter, half of Wall Street expectations. The company’s market capitalization hovered near $1 trillion as its shares soared 30% the next day. Nvidia CEO Jensen Huang announced on May 29 that the world is at “a turning point in a new era of computing.”
other chip companies, from other designers such as supermicro to the manufacturer, e.g. TSMC Taiwan, has swept artificial intelligence Excited. The same goes for suppliers of other computing infrastructure — everything from those colorful cables, noisy air-conditioning units and data center floor space to the software that helps run computers. artificial intelligence Model and marshal data. An equal-weighted index of more than 30 such companies is up 40% since Chatcommon technologyLaunched in November vs. 13% for tech-heavy companies Nasdaq Index (see chart). “A new technology stack is emerging,” concludes Daniel Jeffries. artificial intelligence Infrastructure Coalition, a lobby group.
On the face of it, artificial intelligence Gubbins seem far less exciting than the clever “big language model” behind Chatcommon technology and its rapidly expanding array of competitors.But as model builders and app makers relying on these models vie for a piece of the future artificial intelligence Pie, they all require computing power in the here and now — and a lot of it.
Newest artificial intelligence Systems that include generative sorting are much more computationally intensive than older systems, not to mention non-artificial intelligence application. Amin Vahdat, Principal artificial intelligence Infrastructure for Google Cloud Platform, the internet giant’s cloud computing arm, has observed a tenfold increase in model size every year for the past six years. common technology-4, support the latest version of the chat functioncommon technology, analyzing data with about 1trn parameters, more than five times that of its predecessor. As the complexity of models increases, so does the computational demand to train them.
Once trained, artificial intelligences requires less number-crunching power to use in a process called inference. But given the range of applications offered, inference will add up and require a lot of processing power. Microsoft has more than 2,500 customers using Open technology servicesartificial intelligencechatcommon technologyof creators, the software giant owns nearly half. This is ten times the previous quarter.Alphabet, Google’s parent company, has six products with 2 billion or more users worldwide — and plans to enhance them with generative tech artificial intelligence.
Demand for computing power has exploded, and the most obvious winners are chipmakers.Companies such as Nvidia and supermicro Manufacturers get a license fee every time they etch their blueprints onto a silicon wafer, e.g. TSMC Represent end customers, especially the large cloud computing providers powering the most powerful clouds artificial intelligence application. artificial intelligence So a boon to chip designers, as it benefits from more powerful chips (which tend to yield higher profits), and more. Swiss bankA bank estimates that in the next one or two years artificial intelligence Will increase demand for specialized chips called graphics processing units (graphics cards) A reduction of $10 billion to $15 billion.
As a result, Nvidia’s annual data center revenue, which accounts for 56% of its sales, could double. supermicro is launching a new graphics card later this year.Although it is a much smaller player graphics card– Design games larger than Nvidia, on a larger scale artificial intelligence Stacy Rasgon, a broker at Bernstein, said the boom meant the company was poised to benefit from the market “even if it just sucked up the dregs of the market.”Chip design startups focus on artificial intelligence, such as Cerebras and Graphcore, are trying to make a name for themselves. Data provider PitchBook counts about 300 such companies.
Of course, some windfalls will also go to the manufacturer.in april TSMCCC Wei, the owner of the artificial intelligence-related demand”. Investors are more enthusiastic. Nvidia’s shares rose 10 percent after its latest earnings report, adding about $20 billion to its market value. Less obvious beneficiaries include technology that allows more chips to be packed into a single processing unit. Co. Besi is a Dutch company that makes tools that help bond chips together. According to Pierre Ferragu of New Street Research, another analyst firm, the Dutch company controls three-quarters of the high-precision bonds. Its shares have risen by more than half this year.
Swiss bank estimate graphics cardIt accounts for about half of the professional cost artificial intelligence server, while a standard server is only one-tenth. But they’re not the only necessary gear.To work as a computer, the data center’s graphics cardWe also need to talk to each other.
This in turn requires increasingly advanced networking equipment such as switches, routers and specialized chips. The market for such kits is expected to grow 40 percent a year over the next few years, reaching nearly $9 billion by 2027, according to research firm 650 Group. Nvidia also licenses such kits, accounting for 78% of global sales. But rivals such as Arista Networks, a California company, are also starting to pay attention to investors: its shares have risen nearly 70% in the past year. Broadcom, which sells specialized chips that help networks operate, says its annual sales of such semiconductors will quadruple to $800 million in 2023.
this artificial intelligence The boom is also good news for the companies that assemble the servers that go into data centers, Peter Rutten points out data center, another research firm.Another analyst firm, Dell’Oro Group, predicts that global data centers will increase artificial intelligence From less than 10% today to about 20% in five years, and the suite’s share of data center server capex will increase from about 20% today to 45%.
This will benefit server makers such as Taiwan’s Wistron and Inventec, which mainly produce custom servers for large cloud providers such as Amazon Web Services (AWS) and Microsoft Azure. Smaller manufacturers should do just fine too. The boss of Wiwynn, another Taiwanese server maker, recently said, artificial intelligence– Related items account for more than half of its current order book.U.S. company Super Micro said that in the three months to April artificial intelligence Products accounted for 29% of its sales, up from an average of 20% in the previous 12 months.
all of these artificial intelligence Hardware requires specialized software to function.Some of these programs come from hardware companies; Nvidia’s software platform, called CUDAAllow customers to take full advantage of their graphics cards, for example.Apps created by other companies allow artificial intelligence Company management data (Datagen, Pinecone, Scale artificial intelligence) Or host large language models (HuggingFace, Replicate). PitchBook counts about 80 such startups. So far this year, more than 20 companies have raised new funds; Pinecone investors include Andreessen Horowitz and Tiger Global, two major venture capital giants.
As with hardware, the primary customers for much of this software are cloud giants. Amazon, Alphabet and Microsoft plan to collectively commit about $120 billion in capital expenditures this year, up from $78 billion in 2022. Much of that will be used to expand its cloud capacity.Even so, demand artificial intelligence Computing power is so high that they’re all struggling to keep up.
This creates an opportunity for challengers.In the past few years IBMNvidia and Equinix are already offering support for graphics cardis “as a service”. artificial intelligenceCloud-focused startups are also proliferating. One of them, Lambda, raised $44 million from investors such as Gradient Ventures, one of Google’s venture capital arms, and Open co-founder Greg Brockman.artificial intelligence. The deal values the company at approximately $200 million. A similar company, CoreWeave, raised $221 million in April, including funding from Nvidia, at a $2 billion valuation. According to Brannin McBee, co-founder of CoreWeave, focusing on customer service and artificial intelligence Help it compete with the cloud giants.
last group artificial intelligence– The infrastructure winner closest to offering an actual shovel: data center landlords. Their properties are being filled as demand for cloud computing explodes. In the second half of 2022, the data center vacancy rate is 3%, a record low. Specialists such as Equinix or its rival Digital Realty are increasingly competing with big asset managers keen to add data centers to their real estate portfolios. In 2021, private market giant Blackstone spent $10 billion to acquire quality control system Realty Trust, one of the largest data center operators in the United States. In April, Blackstone’s Canadian rival Brookfield acquired French data center company Data4, which has been investing heavily in data centers.
continued growth of artificial intelligence– The infrastructure stack may experience limitations. One is energy. A big investor in data centers points out that the availability of electricity, of which data centers are huge users, is expected to slow the development of new data centers in hubs like Northern Virginia and Silicon Valley.Another potential hurdle is the transition from giant artificial intelligence Models and cloud-based inferencing reason about smaller systems that require less computing power to train and can run inference on smartphones, like Google’s recently scaled-down PALM Model.
The biggest question mark is the durability of the AI boom itself.Although chat is popularcommon technology And its ilk, a profitable use case for the technology remains unclear. In Silicon Valley, hype can turn into disappointment in an instant. Nvidia’s market cap soars in 2021 as its graphics cardIt turned out to be great for mining bitcoin and other cryptocurrencies, but collapsed as the cryptocurrency boom turned to bust.
If the technology does live up to its transformative toll, regulators could crack down.Policymakers around the world worry about generating artificial intelligenceThe possibility of eliminating jobs or spreading misinformation, safeguards are already being considered.In fact, on May 11 lawmakers in European Union A set of rules was proposed to limit chatbots.
All these constraints may slow down artificial intelligenceof the deployment, and in doing so suppressed the foreground artificial intelligence– Infrastructure companies. But maybe only a little bit.Even if generated artificial intelligence While it’s not quite as transformative as its proponents claim, it will almost certainly be more useful than encryption.and many others, non-generated artificial intelligenceThis also requires a lot of computing power.global ban on generation artificial intelligence, which is not on the horizon, may well stop the gold rush. As long as everyone is running around, the pick and shovel peddlers will make a fortune. ■
To stay on top of the most important news in business and technology, subscribe to The Bottom Line, our weekly newsletter for subscribers.