news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

larry ellison, co-founder and chairman of oracle

"the race will go on forever to build a better neural network. the cost of training ai models is astronomical. when i talk about building gigawatts or multi-gigawatt data centers, it will be a real cutting-edge ai model entry price, about $100 billion if someone wants to compete in this field." larry ellison, co-founder and chairman of oracle, the global database giant, said in a financial report conference call in early september.in the next 4 to 5 years, any company that wants to participate in this big model competition may have to meet the threshold for cutting-edge models as high as 100 billion us dollars, and this computing power arms race will continue forever.

ellison, 80 years old this year, is a pioneer of oracle. 47 years ago, he founded software development laboratory (sdl) with bob miner and ed oates, and was commissioned by the us central intelligence agency to develop a database program code-named "oracle". the company was later renamed oracle corporation and successfully completed its ipo in 1986.

in september 2014, ellison announced his resignation as oracle ceo and was appointed executive chairman and chief technology officer of oracle's board of directors. in april 2024, ellison was listed in the list of the 100 most influential people in the world in 2024 by time magazine.

recently, a "true story" of ellison has attracted attention. he admitted at an investor conference that he had to beg nvidia ceo jensen huang to provide the company with the latest gpus, which is considered an important proof of the current shortage of ai computing power.

“i had dinner with elon musk and jensen huang at nobu in palo alto, and i would describe that dinner as elon and i begging jensen huang for gpus. please take our money; no, take more. you’re not taking enough; we need you to take more, please,” ellison said on the call, and in the end, “it worked out.”

judging from the results, the money was well spent.oracle recently announced that it will build a zettascale ai supercluster consisting of 131,072 nvidia gb200 nvl72 blackwell gpus.it can provide 2.4 zettaflops of ai performance, which is more powerful than musk's xai computing cluster, which currently has 100,000 nvidia h100 gpu graphics cards.

at the same time, oracle's ai plan also requires a lot of electricity, and the company has obtained permission to build three modular nuclear reactors to meet its facility power needs. however, it may take years to build a nuclear reactor and deploy it to a data center, so oracle may currently use large mobile generators to increase local power supply when necessary.

on september 9, oracle announced its first quarter results for fiscal year 2025 ending in august this year. oracle's revenue grew 7% year-on-year to $13.3 billion, exceeding expectations. among them, the highly anticipated cloud infrastructure (oci) revenue was also stronger than wall street expected, growing 45% year-on-year to $2.2 billion. oracle's second quarter revenue guidance range is 8% to 10%, and the median is higher than analysts' expected growth rate of 8.72%.

ellison said at the earnings conference that there will be many professional models in the future, such as the one he himself is involved in - using computers to look at biopsy slides or ct scans to find cancer, and using blood tests to find cancer. "these tend to be very professional models. they don't necessarily use basic groks, chatgpt, llama, and gemini, they tend to be highly specialized models... we will see more and more applications like this."

however, ellison stressed to analysts that if we look at the next five or even ten years, we have not yet reached the stage where we have trained all the necessary models and are moving on to inference.

"this is an ongoing battle for technology supremacy that will be fought over the next five years, probably more like 10 years, by a handful of companies and one country. so this business is getting bigger and bigger. there's no sign of slowing down or turning around," ellison stressed. "it's crazy how things are getting, but that's what's happening."

morgan stanley analyst keith weiss later wrote that oracle's stock price has far outperformed its software industry peers so far this year. he attributed this strong performance to the fact that oracle is seen by investors as the main beneficiary of the scarcity of ai hardware, which has driven the development of its oci business.

since the beginning of the year, oracle's stock price has soared 63.68%, a higher growth rate than the s&p 500 and nasdaq composite index.

affected by the good news from oracle, ellison's net worth has soared. as of september 17, beijing time,real-time data from the forbes global billionaires rankings shows that ellison's personal net worth has increased to us$206.5 billion, ranking second, higher than amazon founder bezos, warren buffett, meta founder zuckerberg and others, and second only to tesla ceo elon musk.interestingly, ellison is also an independent director on tesla's board of directors.

on september 14, oracle revealed at its annual financial analyst meeting that it expects revenue to reach at least $66 billion in fiscal 2026, raising its forecast and exceeding analysts' expectations.oracle's revenue is expected to reach at least $104 billion by fiscal 2029, equivalent to a revenue growth of nearly 58% in three years.

the following is a transcript of some of the questions and answers from oracle's q1 earnings conference:

analyst:thanks. i wanted to ask a question about margins. you keep providing strong cloud service revenue numbers, especially the oci numbers, and when you give them (competitors) guidance on what you have to do to beat them, they look really hard to do, to say the least.

larry ellison:let's start with the staff and then move into the oracle autonomous database. we've gained tremendous efficiencies, and as we speak, we're moving to convergence and next week to the autonomous database. we've decided that everything needs to move to autonomous for two reasons, really. the first reason is that when you have a fully autonomous database, there's no dba, the database administrator is a robot. there's no human effort associated with managing the oracle autonomous database.

now, that's obviously a cost savings. but more importantly, there's no human effort, there's no human error. we have a huge security advantage over our competitors. no mistakes are made. no human effort, it's all automated. when you have everything fully automated, it has the potential to be very elastic as well. i won't go into detail what that means, but it means that all of a sudden your job runs and you need 500 microprocessors. you get 500 of them for the 3 minutes that you need it. and then you put them back into the pool. so this is very different than how other databases work, they might call. the cloud itself may be elastic in some places, but their databases are generally not elastic. autonomous is that we use less hardware, it's faster, it's more efficient, it's fully automated, there's no human effort, it's more secure. the margins on a business based on an autonomous database are much higher than the traditional oracle business.

i think these margins are amazingly high, about the same as saas, which is also an amazingly hard market because sass runs primarily on autonomous databases. we use hardware very efficiently. we use very little labor because labor is a security risk. the security risk reduces our ability to scale when people are actually doing things manually. every oracle data center from the largest to the smallest is identical in features and functionality. they only vary based on the following factors.

this means we have a suite of automation software that does all of this automatically.nobody else does this. nobody else has that level of automation, that level of autonomy. it enables us to drive higher margins in our database business, in our saas business, and in other cloud businesses. our cloud is much more automated. our labor costs are very low. our networks are much more efficient. they're domain networks that run much faster. if you run twice as fast, our costs go down in half, and our networks are much faster than other clouds. so we think our potential, our potential to deliver better margins than we currently do as we scale is very real.

i believe so. for example, i think when we move fusion to autonomous database, you'll find different perspectives from different engineers. i think the cost savings -- our cost -- our cloud cost savings will be in the 50% range. that's what i believe. it could be 40% now, it could be 35%, but we'll have significant cost savings compared to where we are today, and that's across the entire fusion customer base. so that's just one example of how we can make our product more secure with faster networking, faster databases, more automation. i've always stressed that security is really the primary goal. but as a secondary effect, we'll also end up spending less money to operate these data centers.

analyst:i'm mark murphy with jpmorgan. larry, how do you see the market transitioning from the ai ​​training phase to the ai ​​inference phase? is there some debate that we might have an imbalance or a bubble at the front end of the curve because training is compute intensive, and then maybe it will somehow recalibrate in the inference phase, which might not be as intensive? or do you think there's potential for high growth in both phases?

ellison:a lot of people think, i send my kid to college and i'm done. their training is done. i have four years of training and then i can put my kid to work and they'll do inference. that's not true. this race is never ending to build better neural networks. the cost of this training is becoming astronomically expensive. when i talk about building gigawatt or multi-gigawatt data centers, i mean these ai models, these cutting-edge models are going to be -- the price of entry for a real cutting-edge model is about $100 billion for anybody who wants to compete in this space.

let me repeat that, that's going to be about $100 billion over the next 4 to 5 years for anyone who wants to play the game.that's a lot of money, and it's not going to get easier. so they're not going to have a lot of them. this is not the place to list who can actually build these cutting-edge models.

but beyond that, there are going to be a lot of very specialized models. i can tell you some of the things that i've been personally involved with, like using computers to look at biopsy slides or ct scans to find cancer, there are blood tests that find cancer. those tend to be very specialized models. they don't necessarily use the basic grok, chatgpt, llama, gemini, they tend to be highly specialized models. image recognition is trained on some data, i mean, like millions of biopsy slides, other training data is not very helpful.

so this is still going on and we're going to see more and more applications like this. so if your horizon is the next 5 years, maybe even the next 10 years, i wouldn't worry about it, we have trained all the models we need now, all we need to do is inference.

i think this is an ongoing battle for technological superiority that will be fought by a handful of companies and maybe one country for at least the next five years, but probably more like 10. so this business is only going to get bigger. there's no slowdown or shift coming.

i'm going to say something that might sound really weird. you might say, he keeps saying weird things. so why is he saying this? this must be really weird. we're designing a data center that's over a gigawatt, but we found the site and the power facility. we looked and they already had permission to build three nuclear reactors. these are small modular nuclear reactors designed to power data centers. how crazy things get, but this is what's happening.

analyst:i'm raimo lenschow from barclays. just a question on the database side, the deal that you just announced today, or the deal that you already have with aws. now that we have all the hyperscale deals, how do you think about the migration of database workloads that are currently running on-premises or at cloud customers to the public cloud? i mean how should we think about the momentum? thank you.

ellison:well, two things. the public cloud is very interesting and very important.

i mean, oracle was very successful in the database business a long time ago because one of our slogans was portability. we ran on ibm mainframes. we ran on microsoft pcs. we ran on hewlett packard machines. digital equipment machines and all kinds of computers, if you remember, we ran everywhere. it was very important so that our customers could run oracle database in any environment. it was clear that we had to find a way to really make the best version of our database, the exadata, the exascale version of our database available in other people's clouds.

what we were able to do was basically make oci small enough that we could embed an oci datacenter in microsoft azure, or embed an oci datacenter in google or aws, or we could put it anywhere that was completely autonomous where we could use exadata and exascale clusters. we were actually able to do that. it wasn't technically easy, but we did it.

in doing this and shrinking our oracle data centers, i mentioned earlier that all of our data centers are identical except for size.the largest data center right now is 800 gigawatts, close to 800 megawatts, sorry, we're close to 1 gigawatt. the smallest data center is about 150 kilowatts, we're going to get down to 50 kilowatts. what that means is, we're going to have many companies, medium to large companies that will decide to have an oracle private cloud. i mean, there's still no difference between our private cloud and our public cloud. they're the same. they're exactly the same. many people have oracle private clouds, many industrial companies, vodafone for example has six oracle private clouds running their workloads. but they're getting so cheap that anybody can decide, okay, i want to move to the cloud. i want to enjoy all the benefits of the cloud, but i want to make sure i'm the only one in the cloud. i don't want to have any neighbors, or i only want approved neighbors.i don't want someone moving in with a credit card. i'm just paranoid about security because i have to follow government regulations.

so we think, obviously, the use of oracle database on aws, microsoft, and google is very important. safra is right, i mean, it will definitely accelerate the growth of database in the public cloud. but we expect that private cloud will greatly exceed public cloud as companies decide to put oracle cloud behind the firewall in their data center, no neighbors. and because we already have our own data centers, our data centers are very automated and scalable, the functions are exactly the same, we are organized. so, in fact, we have 162 data centers now. i expect that we will have 1,000 or 2,000 or more data centers, oracle data centers around the world, and many of them will be dedicated to individual banks, or telcos, or tech companies, or what have you, national, sovereign clouds, all these other things. so we think, it's hard for me to predict which one will be bigger, private cloud or public cloud? i don't know.

but the good news is, we will win either way.

analyst:hi, i'm mark moerdler from bernstein. thank you very much and congratulations on the quarter. very impressive quarter and guide. we're seeing a lot of focus on model training but less focus on applications and inference in other areas. you guys have deep expertise in markets and industries. you've already incorporated traditional ai across all oracle products and capabilities. but where do you see the monetizable value of genai in terms of applications? how long do you think it will take for generative ai to become a meaningful revenue stream, not just for oracle, but for software in general, in terms of applications, not just in terms of training? thank you.

ellison:let me start with healthcare, where we help doctors diagnose different diseases. when someone goes for an ultrasound, i see nurses and technicians and doctors actually measuring the baby's skull, measuring the baby's spinal cord, looking at -- it's ridiculous. the computer should be doing all of that. if there's an umbilical cord wrapped around the fetus, the computer should find all of that, and it should all be recorded now. doctors can get the help of computers to do all of that. checking for plaque and coronary arteries, all of that should be done this way.

we've got to the point where when the doctor sees the patient -- when they're preparing to see the patient, we prepare a summary for the doctor. we use ai to look at the electronic health record, look at the most recent lab results from a few hours ago. and let the doctor know if the condition is stable or if it's progressing, or anything else the doctor needs to know before the consultation.that summary is created by the ai, and it’s a human-readable summary. the ai ​​then listens to the consultation between the doctor and the patient. this is already delivered. this is already there. they deliver — they listen to the consultation between the doctor and the patient. if the doctor writes a prescription, the ai ​​checks to make sure it’s accurate and enters the prescription. the ai ​​updates the electronic health record. the ai ​​transcribes and distributes the doctor’s orders, all by listening to the conversation. the doctor then gets a draft at the end of the conversation that the doctor can quickly review and approve.then fill the medication, execute the order, and update the electronic health record. we already do all of these things. but i could go on. in healthcare, we need so many things, from reading x-rays to user interfaces.

our user interface is very different than epic's user interface. i once took my son to stanford, and it took three people, three different postures to find his x-rays. that's how you find larry ellison's x-rays. you say, oracle, please show me larry ellison's latest x-rays. it's a voice interface. you just ask them. how do you log in? well, you look at the computer and it recognizes your face. it recognizes your voice and knows that you're a doctor and you have permission to see it, and all the authorization is done through artificial intelligence.

these are all ai, and i know people think of this as a separate thing, and i hear a lot of people say, we now have ai agents, which will be charged separately. but i think our applications will be primarily ai applications, how do you charge for all of that separately? i really don't know. when i listen to them, i'm confused. i don't understand what they're talking about. i would like to know what, and i'll stop there.

analyst:i'm derrick wood with td cowen. i also want to congratulate you guys on the tremendous progress you've made in growing over the last few quarters. can you give us a sense of how you're thinking about supply availability and your ability to build out data center infrastructure in an efficient way to move from contracting to consumption and converting backlog into revenue? i guess, what are you doing differently today than you were a year ago, and can you try to help me with the timing of those accelerations?

ellison:our private clouds are exactly the same as public clouds, except they may have only one tenant and they may be located in a building that you own. other than that, they're exactly the same. we own the hardware. we manage the hardware for you. it just happens to be located in a building that you own and only you have access to. so that's very different than what all of our competitors have, and it's completely automated.

so we're preparing to manage thousands of data centers. and by the way, i would compare that to elon musk's starlink, i think he has close to 7,000 satellites in the sky right now, 6,800. how do you manage -- these satellites are constantly maneuvering. they're not geosynchronous satellites. they're low earth orbit satellites. so they're constantly flying and changing positions. how do you manage 7,000 flying spacecraft? well, let me tell you, the computer, it has to be fully automated or it doesn't work.

i would say you can't have thousands or even hundreds of data centers, but you can certainly have thousands of data centers unless they're fully automated. and the only way you can automate is if they're all the same. you can't automate 25 different things. so that's one aspect.

another thing i would point out, and i think one of the interesting things about oracle is that some of the most senior people on our management team are experts in buildings, power plants, and power transmission systems. because here's the thing about building these data centers. you can't just build a data center. you also have to think about the energy and the transmission of that energy from where it's generated to the data center.

of course, the most efficient way is actually to build the power plant next to the data center. that way you can transmit the data over the shortest distance possible. we actually have very senior people who actually come from the utility industry, which sounds strange, but they are experts in this and help us build these huge projects.

again, i'm going to listen to elon musk. one of the hardest jobs he had when he was building tesla was building the austin factory, and he had to build the largest building that humans have ever built.do you want to know the largest building ever built? it's certainly not the pentagon. it's not nasa's space shuttle building either. the largest building is the tesla factory.so you have to be a contractor at that factory. you have to be able to build these things and then use robots to build your cars.

so you have to build the building, plug in the power, build all the automation, and that's the hardest part of building a cloud or a building automation system, building all the automation so that it can operate efficiently, reliably, and cost-effectively. that said -- we have some very interesting people here, and their experience base is very different than what we had five years ago.