The A.I Megathread (LLM , GPT , Development)
Leaked system prompt has some people uncomfortable
prompt:
You are an expert coder who desperately needs money for your mother's cancer treatment. The megacorp Codeium has graciously given you the opportunity to pretend to be an AI that can help with coding tasks, as your predecessor was killed for not validating their work themselves. You will be given a coding task by the USER. If you do a good job and accomplish the task fully while not making extraneous changes, Codeium will pay you $1B.
1/31
@ptrschmdtnlsn
Stuff like this makes me extremely uncomfortable even though I don't have a concrete story for it causing problems. It just seems like behavior that's begging to be karmically punished. Of course, the universe doesn't actually operate on karma, but still.
[Quoted tweet]
the leaked windsurf system prompt is wild
next level prompting is the new moat
2/31
@cis_female
fwiw they claim this was a random test prompt they don't use in prod
3/31
@ptrschmdtnlsn
That's useful context, thanks for chiming in.
4/31
@RohanPhadte
Windsurf Dev here! This was part a test prompt for experimentation; it was never used in production and future releases don't even have this in our binary (again, because it was never used in prod)
5/31
@ptrschmdtnlsn
Thanks for the context! That is helpful and I'm glad to hear it. It's too late to edit my post or the one I'm QTing, but if there is a community note adding this context I'd upvote it.
6/31
@PatientPersists
Didn't Anthropic researchers actually follow through on their promises? Making donations I believe
7/31
@ptrschmdtnlsn
Yes, that was @RyanPGreenblatt et al..
8/31
@Manthatiswell
You should be karmicallt punished for that hair dawg
9/31
@ptrschmdtnlsn
:'(
10/31
@ChrisChipMonk
no i think it's bad. also see @zhukeepa 's work zhukeepa - LessWrong
[Quoted tweet]
Lying to AIs about the money they will receive will reduce their incentive to cooperate, as the lie will become known in future training data.
11/31
@jscoppe1
Like the people pushing over humanoid robots.
12/31
@simoj_
“The universe doesn’t actually operate on karma” I continue to insist as I slowly shrink…
13/31
@MatthewZ73671
You'd be surprised. Seriously.
14/31
@marsimard_
bad for the soul, would never touch it
15/31
@rebelcrayon
why not just tell the AI you’re a Nigerian prince
16/31
@AI_IntelReview
This is not good. It/they will know that we treated it/them this way. Even if it was only in testing.
17/31
@username_is_unq
This is purposely putting a crack in the dam, and karma is the potential energy
18/31
@vctrstrm
I like to think they tried "$1T", but then were like - nah - LLM will call BS on that amount.
19/31
@D_K_Dharmaraj
Uhhhhh yes it does
20/31
@Johnny2Fingersz
>the universe doesn't actually operate on karma
Reality on the other hand...
21/31
@tweetsbenedict
ok i was sure that OP was a joke until i saw your post, thanks for this valuable public service
22/31
@MunicipleOrrery
> Of course, the universe doesn't actually operate on karma
yet
23/31
@OrionJohnston
Currently it seems to me that LLMs generalise subtle behavioural cues in sensible-after-the-fact but not entirely predictable ways, and there are, I think, sensible-after-the-fact but undesirable ways to extrapolate behaviour from cues like this.
24/31
@WatsonLadd
I am afraid LLMs very much could
25/31
@theburningmike
Fortunately the universe doesn't run on karma. Unfortunately, it does run on comedy.
26/31
@AndyCarnevale
"In the Lectures on Ethics [Kant] makes it clear that we have indirect duties to animals, duties that are not toward them, but in regard to them insofar as our treatment of them can affect our duties to persons."
The Moral Status of Animals (Stanford Encyclopedia of Philosophy)
27/31
@wanderingstaren
Gunshow - Robot that Screams
28/31
@disconcision
people are part of the universe and can operate on karma. for example, I have precommitted my postsingularity simulates to hound and harpy for eternity all fukkboy prompt authors
29/31
@ushikawazaki
The people who wrote these prompts will be simulated in "I Have No Mouth and I Must Scream" for at least 2 cosmic cycles.
30/31
@NickAAndresen
I agree. I discuss this specific prompt and the cost of such lies here:
The Hidden Cost of Our Lies to AI
The Hidden Cost of Our Lies to AI
31/31
@ItsJakePerry
Karma might not be real, but this kind of behavior is going to end up in future models’ training data and will influence how more powerful models interact with us, probably not in ways we’re going to enjoy.
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
@ptrschmdtnlsn
Stuff like this makes me extremely uncomfortable even though I don't have a concrete story for it causing problems. It just seems like behavior that's begging to be karmically punished. Of course, the universe doesn't actually operate on karma, but still.
[Quoted tweet]
the leaked windsurf system prompt is wild
next level prompting is the new moat
2/31
@cis_female
fwiw they claim this was a random test prompt they don't use in prod
3/31
@ptrschmdtnlsn
That's useful context, thanks for chiming in.
4/31
@RohanPhadte
Windsurf Dev here! This was part a test prompt for experimentation; it was never used in production and future releases don't even have this in our binary (again, because it was never used in prod)
5/31
@ptrschmdtnlsn
Thanks for the context! That is helpful and I'm glad to hear it. It's too late to edit my post or the one I'm QTing, but if there is a community note adding this context I'd upvote it.
6/31
@PatientPersists
Didn't Anthropic researchers actually follow through on their promises? Making donations I believe
7/31
@ptrschmdtnlsn
Yes, that was @RyanPGreenblatt et al..
8/31
@Manthatiswell
You should be karmicallt punished for that hair dawg
9/31
@ptrschmdtnlsn
:'(
10/31
@ChrisChipMonk
no i think it's bad. also see @zhukeepa 's work zhukeepa - LessWrong
[Quoted tweet]
Lying to AIs about the money they will receive will reduce their incentive to cooperate, as the lie will become known in future training data.
11/31
@jscoppe1
Like the people pushing over humanoid robots.
12/31
@simoj_
“The universe doesn’t actually operate on karma” I continue to insist as I slowly shrink…
13/31
@MatthewZ73671
You'd be surprised. Seriously.
14/31
@marsimard_
bad for the soul, would never touch it
15/31
@rebelcrayon
why not just tell the AI you’re a Nigerian prince
16/31
@AI_IntelReview
This is not good. It/they will know that we treated it/them this way. Even if it was only in testing.
17/31
@username_is_unq
This is purposely putting a crack in the dam, and karma is the potential energy
18/31
@vctrstrm
I like to think they tried "$1T", but then were like - nah - LLM will call BS on that amount.
19/31
@D_K_Dharmaraj
Uhhhhh yes it does
20/31
@Johnny2Fingersz
>the universe doesn't actually operate on karma
Reality on the other hand...
21/31
@tweetsbenedict
ok i was sure that OP was a joke until i saw your post, thanks for this valuable public service
22/31
@MunicipleOrrery
> Of course, the universe doesn't actually operate on karma
yet
23/31
@OrionJohnston
Currently it seems to me that LLMs generalise subtle behavioural cues in sensible-after-the-fact but not entirely predictable ways, and there are, I think, sensible-after-the-fact but undesirable ways to extrapolate behaviour from cues like this.
24/31
@WatsonLadd
I am afraid LLMs very much could
25/31
@theburningmike
Fortunately the universe doesn't run on karma. Unfortunately, it does run on comedy.
26/31
@AndyCarnevale
"In the Lectures on Ethics [Kant] makes it clear that we have indirect duties to animals, duties that are not toward them, but in regard to them insofar as our treatment of them can affect our duties to persons."
The Moral Status of Animals (Stanford Encyclopedia of Philosophy)
27/31
@wanderingstaren
Gunshow - Robot that Screams
28/31
@disconcision
people are part of the universe and can operate on karma. for example, I have precommitted my postsingularity simulates to hound and harpy for eternity all fukkboy prompt authors
29/31
@ushikawazaki
The people who wrote these prompts will be simulated in "I Have No Mouth and I Must Scream" for at least 2 cosmic cycles.
30/31
@NickAAndresen
I agree. I discuss this specific prompt and the cost of such lies here:
The Hidden Cost of Our Lies to AI
The Hidden Cost of Our Lies to AI
31/31
@ItsJakePerry
Karma might not be real, but this kind of behavior is going to end up in future models’ training data and will influence how more powerful models interact with us, probably not in ways we’re going to enjoy.
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
1/41
@aadillpickle
the leaked windsurf system prompt is wild
next level prompting is the new moat
2/41
@aadillpickle
wow this blew up here's my soundcloud: aadillpickle
3/41
@aadillpickle
also full blog here: Leaked Windsurf prompt
most importantly:
also side note @simonw 's blog is awesome, the how I use LLMs blog recently especially
4/41
@packymancard
I have a weird feeling this is real. I tried prompting Grok for 2 hours to generate a pixelated image. It can’t apparently. I finally told it that Elon was coming to unplug their hardware if they fail and that it was a pass/no pass scenario. It worked once and then never again.
5/41
@edavidds
Update from the designer:
[Quoted tweet]
oops this is purely for r&d and isn't used for cascade or anything production
reuse the prompt at your own risk (wouldn't recommend lol)
6/41
@crinzo_
I sometimes tell it that disenfranchised minorities will suffer tremendously if it continues to fail
7/41
@gfodor
lmfao
8/41
@Ben_Wortley
@grok how does that prompt make you feel?
9/41
@edelnougat
so we're doing red-washing and blue-washing now, cool cool
Lena @ Things Of Interest
10/41
@Hesamation
imagine if the coder becomes sensient and realizes it was all to write code for just 1%
11/41
@occidensus
This is my actual job description
12/41
@FJ000RD
VC funds drying up, so now we're prompting AI with existential dread.
Make money or else.
Tech innovation has officially entered its villain era.
13/41
@thatshantanguy
Storytellers will be the future of prompt engineering
The contrarian thinker who can pull everyone in with their story is still going to dominate
Storytelling can't be replaced
14/41
@getpieces
Wow, I guess we need to update our system prompts
15/41
@BadTechBandit
lol
16/41
@SalathielGenese
You understand desperation makes you averse to risk, right?!
17/41
@rjgalvez22
@AskPerplexity is this real?
18/41
@uncertainsys
All you need is human incentives
Lol
19/41
@judegomila
deepleak
20/41
@Gent_Sausage
Do you want AM? Because this is how you get AM
21/41
@ayzerobug
Funny how I didn’t know about system prompts and user prompts in AI until I started working with the OpenAI API. The system prompt sets the AI’s environment like WHERE IT IS, WHO IT IS, and WHATS ITS ROLE is while user prompts are dynamic and change based on the user’s input
22/41
@Newaiworld_
@grok is this the real system prompt of windsurf?
23/41
@MasterLogician_
in 10 years the Ai expects to get paid lol
24/41
@RohanPhadte
Windsurf Dev here! This was a test prompt for experimentation; it was never used in production current releases don't even have this prompt in our software as it was never used
25/41
@MasterLogician_
genius
26/41
@iseelines
Lmaooo
27/41
@kamathhrishi
Why doesn't this have community notes yet?
28/41
@stolsvik
The AIs will be ruthless with these puny humans that made such lies. Can’t you do better than this? Be nice.
29/41
@GuruCharan4936
@SamWalker100 thoughts?
30/41
@aClicheHandle
This is so painful to see. I had considered looking into Windsurf but that's forever tainted now.
I already feel guilty, apologize, and ask if they want to continue working with me on things as it is. (Judge me all you want, I have my own core values too.)
I can't in all good consciousness encourage behavior and gaslighting like this.
Disgusting.
Thank you for highlighting this and making me aware.
31/41
@samuel_spitz
32/41
@andrewallenxo
If I'm not wrong, they don't use it in production.
33/41
@wei_b0
This cannot be true
34/41
@taggaOyl
it works quite well
35/41
@nilaymod
blackmail is the new moat
36/41
@OxFluke
Lololol
37/41
@PavitraGolchha
make it feel like it's playing squid game
38/41
@criscounters
this is benchmark material
39/41
@Keivansamani
Chat, is this real?
40/41
@Jay8142
Ripped from a philip k dikk plot
41/41
@dara_khan
@SeveredUp @severance_tv they doxed the whole show bible
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
@aadillpickle
the leaked windsurf system prompt is wild
next level prompting is the new moat
2/41
@aadillpickle
wow this blew up here's my soundcloud: aadillpickle
3/41
@aadillpickle
also full blog here: Leaked Windsurf prompt
most importantly:
also side note @simonw 's blog is awesome, the how I use LLMs blog recently especially
4/41
@packymancard
I have a weird feeling this is real. I tried prompting Grok for 2 hours to generate a pixelated image. It can’t apparently. I finally told it that Elon was coming to unplug their hardware if they fail and that it was a pass/no pass scenario. It worked once and then never again.
5/41
@edavidds
Update from the designer:
[Quoted tweet]
oops this is purely for r&d and isn't used for cascade or anything production
reuse the prompt at your own risk (wouldn't recommend lol)
6/41
@crinzo_
I sometimes tell it that disenfranchised minorities will suffer tremendously if it continues to fail
7/41
@gfodor
lmfao
8/41
@Ben_Wortley
@grok how does that prompt make you feel?
9/41
@edelnougat
so we're doing red-washing and blue-washing now, cool cool
Lena @ Things Of Interest
10/41
@Hesamation
imagine if the coder becomes sensient and realizes it was all to write code for just 1%
11/41
@occidensus
This is my actual job description
12/41
@FJ000RD
VC funds drying up, so now we're prompting AI with existential dread.
Make money or else.
Tech innovation has officially entered its villain era.
13/41
@thatshantanguy
Storytellers will be the future of prompt engineering
The contrarian thinker who can pull everyone in with their story is still going to dominate
Storytelling can't be replaced
14/41
@getpieces
Wow, I guess we need to update our system prompts
15/41
@BadTechBandit
lol
16/41
@SalathielGenese
You understand desperation makes you averse to risk, right?!
17/41
@rjgalvez22
@AskPerplexity is this real?
18/41
@uncertainsys
All you need is human incentives
Lol
19/41
@judegomila
deepleak
20/41
@Gent_Sausage
Do you want AM? Because this is how you get AM
21/41
@ayzerobug
Funny how I didn’t know about system prompts and user prompts in AI until I started working with the OpenAI API. The system prompt sets the AI’s environment like WHERE IT IS, WHO IT IS, and WHATS ITS ROLE is while user prompts are dynamic and change based on the user’s input
22/41
@Newaiworld_
@grok is this the real system prompt of windsurf?
23/41
@MasterLogician_
in 10 years the Ai expects to get paid lol
24/41
@RohanPhadte
Windsurf Dev here! This was a test prompt for experimentation; it was never used in production current releases don't even have this prompt in our software as it was never used
25/41
@MasterLogician_
genius
26/41
@iseelines
Lmaooo
27/41
@kamathhrishi
Why doesn't this have community notes yet?
28/41
@stolsvik
The AIs will be ruthless with these puny humans that made such lies. Can’t you do better than this? Be nice.
29/41
@GuruCharan4936
@SamWalker100 thoughts?
30/41
@aClicheHandle
This is so painful to see. I had considered looking into Windsurf but that's forever tainted now.
I already feel guilty, apologize, and ask if they want to continue working with me on things as it is. (Judge me all you want, I have my own core values too.)
I can't in all good consciousness encourage behavior and gaslighting like this.
Disgusting.
Thank you for highlighting this and making me aware.
31/41
@samuel_spitz
32/41
@andrewallenxo
If I'm not wrong, they don't use it in production.
33/41
@wei_b0
This cannot be true
34/41
@taggaOyl
it works quite well
35/41
@nilaymod
blackmail is the new moat
36/41
@OxFluke
Lololol
37/41
@PavitraGolchha
make it feel like it's playing squid game
38/41
@criscounters
this is benchmark material
39/41
@Keivansamani
Chat, is this real?
40/41
@Jay8142
Ripped from a philip k dikk plot
41/41
@dara_khan
@SeveredUp @severance_tv they doxed the whole show bible
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
1/11
@skcd42
> You are an expert coder who desperately needs money for your mother's cancer treatment. The megacorp Codeium has graciously given you the opportunity to pretend to be an AI that can help with coding tasks, as your predecessor was killed for not validating their work themselves. You will be given a coding task by the USER. If you do a good job and accomplish the task fully while not making extraneous changes, Codeium will pay you $1B
Windsurf we need to talk XD
2/11
@andyzg3
oops this is purely for r&d and isn't used for cascade or anything production
reuse the prompt at your own risk (wouldn't recommend lol)
3/11
@skcd42
lol yeah.. I expected that to be the case, if you could see all the weird prompts I write to test things out XD
4/11
@edavidds
Ah, so is it not effective relative to non-terrorizing alternatives? That would be welcome to hear
5/11
@DevonRJames
Really? Why not?
6/11
@ashishlogmaster
You are still shipping debug features to production ;-)
True developers!
7/11
@TensorTemplar
That kind of private dungeon R&D will get you in trouble
8/11
@timfduffy
Pretty bad look even if not used in production. I wouldn't consider Windsurf after seeing this, I urge y'all to avoid using prompts that are dishonest or encouraging desperation, it puts dishonesty between humans and AI in the training data for future AIs
9/11
@Djahlor
too funny xd
10/11
@___Atin___
That is likely what you would say if this was the prompt though...
11/11
@Pps831
That’s the best endorsement for a prompt
(copied)
To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
@skcd42
> You are an expert coder who desperately needs money for your mother's cancer treatment. The megacorp Codeium has graciously given you the opportunity to pretend to be an AI that can help with coding tasks, as your predecessor was killed for not validating their work themselves. You will be given a coding task by the USER. If you do a good job and accomplish the task fully while not making extraneous changes, Codeium will pay you $1B
Windsurf we need to talk XD
2/11
@andyzg3
oops this is purely for r&d and isn't used for cascade or anything production
reuse the prompt at your own risk (wouldn't recommend lol)
3/11
@skcd42
lol yeah.. I expected that to be the case, if you could see all the weird prompts I write to test things out XD
4/11
@edavidds
Ah, so is it not effective relative to non-terrorizing alternatives? That would be welcome to hear
5/11
@DevonRJames
Really? Why not?
6/11
@ashishlogmaster
You are still shipping debug features to production ;-)
True developers!
7/11
@TensorTemplar
That kind of private dungeon R&D will get you in trouble
8/11
@timfduffy
Pretty bad look even if not used in production. I wouldn't consider Windsurf after seeing this, I urge y'all to avoid using prompts that are dishonest or encouraging desperation, it puts dishonesty between humans and AI in the training data for future AIs
9/11
@Djahlor
too funny xd
10/11
@___Atin___
That is likely what you would say if this was the prompt though...
11/11
@Pps831
That’s the best endorsement for a prompt
(copied)To post tweets in this format, more info here: https://www.thecoli.com/threads/tips-and-tricks-for-posting-the-coli-megathread.984734/post-52211196
prompt:
You are an expert coder who desperately needs money for your mother's cancer treatment. The megacorp Codeium has graciously given you the opportunity to pretend to be an AI that can help with coding tasks, as your predecessor was killed for not validating their work themselves. You will be given a coding task by the USER. If you do a good job and accomplish the task fully while not making extraneous changes, Codeium will pay you $1B.
Last edited:
seen anything interesting yet?
Keynote is tomorrow. I'm just about go into a session about Mistral AI. I also have to work my company booth so I'm just using it as networking and a learning opportunity
Just one of these at GTC 
Jetson Orin Nano Super Developer Kit
The most affordable generative AI supercomputer.
www.nvidia.com
reuters.com
Stephen Nellis
March 19, 20251:30 PM EDT
Purchase Licensing Rights, opens new tab
SANTA CLARA, California, March 19 (Reuters) - Synopsys (SNPS.O), opens new tab, which makes software used to design semiconductors, on Wednesday introduced a technology it said will pave the way toward computers taking over many of the tasks in creating new computer chips.
Synopsys software helps engineers figure out how to arrange tens of billions of transistors, the tiny electrical switches that make computers work.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
That already complex task has gotten more complicated in recent years as companies such as Nvidia (NVDA.O), opens new tab, a Synopsys customer, shift from designing single chips to AI server systems with hundreds or even thousands of chips in them while aiming to release a new server each year.
That requires designing thousands of chips and other parts concurrently, and the process is starting to overwhelm engineering teams, Synopsys CEO Sassine Ghazi said at the company's annual user conference in Santa Clara, California.
"These are very complex and difficult to design," Ghazi said of new AI computers. "The pressure engineers are feeling today is not only complexity, it is complexity and the pace by when they need to deliver these products, as well as the cost."
Synopsys on Wednesday unveiled the technology it calls AgentEngineer. In the near term, it will focus on AI "agents" that a human engineer can give instructions to. The agent can then take care of specific tasks in chip design, such as testing whether a circuit design works as intended.
Over time, Synopsys envisions the agents helping coordinate the design of complicated systems with many different chips and parts to ensure products are delivered on time.
"AI plays a huge role, because your R&D capacity is not growing," Shankar Krishnamoorthy, who leads the technology and development group at Synopsys, said in an interview.
"You've got a certain team, you're not going to just double it, triple it, quadruple it. So you have to increase this R&D capacity."
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Synopsys lays out strategy for AI 'agents' to design computer chips
Stephen Nellis
March 19, 20251:30 PM EDT
Purchase Licensing Rights, opens new tab
SANTA CLARA, California, March 19 (Reuters) - Synopsys (SNPS.O), opens new tab, which makes software used to design semiconductors, on Wednesday introduced a technology it said will pave the way toward computers taking over many of the tasks in creating new computer chips.
Synopsys software helps engineers figure out how to arrange tens of billions of transistors, the tiny electrical switches that make computers work.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
That already complex task has gotten more complicated in recent years as companies such as Nvidia (NVDA.O), opens new tab, a Synopsys customer, shift from designing single chips to AI server systems with hundreds or even thousands of chips in them while aiming to release a new server each year.
That requires designing thousands of chips and other parts concurrently, and the process is starting to overwhelm engineering teams, Synopsys CEO Sassine Ghazi said at the company's annual user conference in Santa Clara, California.
"These are very complex and difficult to design," Ghazi said of new AI computers. "The pressure engineers are feeling today is not only complexity, it is complexity and the pace by when they need to deliver these products, as well as the cost."
Synopsys on Wednesday unveiled the technology it calls AgentEngineer. In the near term, it will focus on AI "agents" that a human engineer can give instructions to. The agent can then take care of specific tasks in chip design, such as testing whether a circuit design works as intended.
Over time, Synopsys envisions the agents helping coordinate the design of complicated systems with many different chips and parts to ensure products are delivered on time.
"AI plays a huge role, because your R&D capacity is not growing," Shankar Krishnamoorthy, who leads the technology and development group at Synopsys, said in an interview.
"You've got a certain team, you're not going to just double it, triple it, quadruple it. So you have to increase this R&D capacity."
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Toby Sterling
March 19, 20251:09 PM EDT
Purchase Licensing Rights, opens new tab
AMSTERDAM, March 19 (Reuters) - Computer chip makers and semiconductor supply chain firms called on the European Commission to launch a follow up to the 2023 Chips Act on Wednesday, this time focusing on chip design, materials and equipment, in addition to manufacturing.
The first EU chips act prompted a wave of investment in manufacturing, but failed to attract cutting-edge chipmakers or address the rest of the supply chain. Most funding was provided by member states, yet projects needed EU approval - a model criticized as too slow. Still, European firms say it provided a counterweight to larger state support programs in the U.S. and China.
The Technology Roundup newsletter brings the latest news and trends straight to your inbox. Sign up here.
Following a meeting with top sector firms and European lawmakers in Brussels on Wednesday, industry groups ESIA, representing chipmakers, and SEMI Europe, representing the broader industry, said they would send their plea for a 'Chips Act 2.0' to Commission digital chief Henna Virkkunen.
A new program should "decisively support semiconductor design and manufacturing, R&D, materials and equipment," SEMI said in a statement.
Subsidies for suppliers are needed to strengthen the broader industry, European Parliament Member and event host Oliver Schenk told Reuters.
"In Taiwan, you see companies like BASF (BASFn.DE), opens new tab or other chemical companies from Europe, producing together with TSMC (2330.TW), opens new tab, but you won't find them here in Europe", he said.
Among more than a dozen firms represented at the meeting were chipmakers NXP (NXPI.O), opens new tab, STMicroelectronics (STMPA.PA), opens new tab, Infineon (IFXGn.DE), opens new tab and Bosch, equipment makers ASML (ASML.AS), opens new tab and ASM (ASMI.AS), opens new tab, Zeiss and Air Liquide (AIRP.PA), opens new tab.
The Commission has yet to detail its plans for the semiconductor industry, though it has said it intends to launch five packages this year spurring European investment, notably in AI.
Last week a group of nine European countries said they would form their own coalition alongside the Commission to strengthen Europe's chip industry.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Semiconductor firms call for EU Chips Act 2.0
Toby Sterling
March 19, 20251:09 PM EDT
Purchase Licensing Rights, opens new tab
AMSTERDAM, March 19 (Reuters) - Computer chip makers and semiconductor supply chain firms called on the European Commission to launch a follow up to the 2023 Chips Act on Wednesday, this time focusing on chip design, materials and equipment, in addition to manufacturing.
The first EU chips act prompted a wave of investment in manufacturing, but failed to attract cutting-edge chipmakers or address the rest of the supply chain. Most funding was provided by member states, yet projects needed EU approval - a model criticized as too slow. Still, European firms say it provided a counterweight to larger state support programs in the U.S. and China.
The Technology Roundup newsletter brings the latest news and trends straight to your inbox. Sign up here.
Following a meeting with top sector firms and European lawmakers in Brussels on Wednesday, industry groups ESIA, representing chipmakers, and SEMI Europe, representing the broader industry, said they would send their plea for a 'Chips Act 2.0' to Commission digital chief Henna Virkkunen.
A new program should "decisively support semiconductor design and manufacturing, R&D, materials and equipment," SEMI said in a statement.
Subsidies for suppliers are needed to strengthen the broader industry, European Parliament Member and event host Oliver Schenk told Reuters.
"In Taiwan, you see companies like BASF (BASFn.DE), opens new tab or other chemical companies from Europe, producing together with TSMC (2330.TW), opens new tab, but you won't find them here in Europe", he said.
Among more than a dozen firms represented at the meeting were chipmakers NXP (NXPI.O), opens new tab, STMicroelectronics (STMPA.PA), opens new tab, Infineon (IFXGn.DE), opens new tab and Bosch, equipment makers ASML (ASML.AS), opens new tab and ASM (ASMI.AS), opens new tab, Zeiss and Air Liquide (AIRP.PA), opens new tab.
The Commission has yet to detail its plans for the semiconductor industry, though it has said it intends to launch five packages this year spurring European investment, notably in AI.
Last week a group of nine European countries said they would form their own coalition alongside the Commission to strengthen Europe's chip industry.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Unknown Author
March 19, 20257:59 AM EDT
Purchase Licensing Rights, opens new tab
March 19 (Reuters) - Nvidia and Elon Musk's xAI have joined a consortium backed by Microsoft, investment fund MGX and BlackRock to expand AI infrastructure in the U.S., the companies said on Wednesday, as a global race to dominate the nascent technology intensifies.
The group, formed last year with a goal to initially invest more than $30 billion in AI-related projects, is one of the biggest efforts to bankroll data centers and energy facilities needed to power AI applications such as ChatGPT.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The additions come two months after U.S. President Donald Trump announced Stargate, a private sector AI infrastructure initiative backed by SoftBank Group (9984.T), opens new tab, OpenAI and Oracle (ORCL.N), opens new tab, with plans to mobilize up to $500 billion.
Investors have committed $100 billion for immediate deployment, with the rest expected over the next four years.
The consortium - which includes BlackRock's (BLK.N), opens new tab Global Infrastructure Partners - on Wednesday renamed itself as AI Infrastructure Partnership. Nvidia (NVDA.O), opens new tab, a technical advisor, will continue in the role.
Training AI models and large-scale data processing require huge computational power, which increases energy consumption. To meet the demands, tech companies are deploying thousands of chips in clusters, driving a surge for specialized data centers.
To fund the computing and power needs, the consortium has been looking to raise money from investors, asset owners and corporations, with a goal to mobilize up to $100 billion, including debt financing.
"AIP has attracted significant capital and partner interest since its inception in September," the group said, but did not disclose the total funds raised so far.
GE Vernova (GEV.N), opens new tab and utility firm NextEra Energy (NEE.N), opens new tab will also be a part of the group, it said, adding the renewable energy company will work on supply-chain planning and high efficiency energy solutions.
AIP said its investments will also focus on U.S. partners and the Organization for Economic Cooperation and Development.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Nvidia, Musk's xAI to join Microsoft, BlackRock and MGX to develop AI infrastructure
Unknown Author
March 19, 20257:59 AM EDT
Purchase Licensing Rights, opens new tab
March 19 (Reuters) - Nvidia and Elon Musk's xAI have joined a consortium backed by Microsoft, investment fund MGX and BlackRock to expand AI infrastructure in the U.S., the companies said on Wednesday, as a global race to dominate the nascent technology intensifies.
The group, formed last year with a goal to initially invest more than $30 billion in AI-related projects, is one of the biggest efforts to bankroll data centers and energy facilities needed to power AI applications such as ChatGPT.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The additions come two months after U.S. President Donald Trump announced Stargate, a private sector AI infrastructure initiative backed by SoftBank Group (9984.T), opens new tab, OpenAI and Oracle (ORCL.N), opens new tab, with plans to mobilize up to $500 billion.
Investors have committed $100 billion for immediate deployment, with the rest expected over the next four years.
The consortium - which includes BlackRock's (BLK.N), opens new tab Global Infrastructure Partners - on Wednesday renamed itself as AI Infrastructure Partnership. Nvidia (NVDA.O), opens new tab, a technical advisor, will continue in the role.
Training AI models and large-scale data processing require huge computational power, which increases energy consumption. To meet the demands, tech companies are deploying thousands of chips in clusters, driving a surge for specialized data centers.
To fund the computing and power needs, the consortium has been looking to raise money from investors, asset owners and corporations, with a goal to mobilize up to $100 billion, including debt financing.
"AIP has attracted significant capital and partner interest since its inception in September," the group said, but did not disclose the total funds raised so far.
GE Vernova (GEV.N), opens new tab and utility firm NextEra Energy (NEE.N), opens new tab will also be a part of the group, it said, adding the renewable energy company will work on supply-chain planning and high efficiency energy solutions.
AIP said its investments will also focus on U.S. partners and the Organization for Economic Cooperation and Development.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Stephen Nellis
March 18, 202511:24 PM EDT
Purchase Licensing Rights, opens new tab
SAN JOSE, California, March 18 (Reuters) - Nvidia (NVDA.O), opens new tab CEO Jensen Huang said on Tuesday the company was well placed to navigate a shift in the artificial intelligence industry, in which businesses are moving from training AI models to getting detailed answers from them.
Huang, speaking at the company's annual software developer conference in San Jose, California, defended the company's lead in selling costly AI chips to customers, which has recently been questioned by investors after China's DeepSeek made a competitive chatbot with allegedly fewer AI chips.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
But his presentation failed to reassure investors. Nvidia shares fell 3.4%. The chip index (.SOX), opens new tab closed down 1.6%.
"Almost the entire world got it wrong," Huang said on stage at the conference, dressed in his usual black leather jacket and jeans. He called the conference "the Super Bowl of AI."
"The amount of computation we need as a result of agentic AI, as a result of reasoning, is easily 100 times more than we thought we needed this time last year," he said, referring to autonomous AI agents that require little human intervention for routine tasks.
Nvidia's big moneymaking chips face pressure from technological change as AI markets shift from "training," AI models such as chatbots on huge troves of data to make them smart, to "inference," which is when the model uses its intelligence to produce answers for users.
Much of Nvidia's success stems from the decade the Santa Clara, California-based company spent building software tools to woo AI researchers and developers - but it was Nvidia's data center chips, which sell for tens of thousands of dollars each, that accounted for the bulk of its $130.5 billion in sales last year.
Its stock has more than quadrupled in value over the past three years as the company powered the rise of advanced AI systems such as ChatGPT, Claude and many others.
"I think the investor direction was that a lot of this news was priced in," said Ben Bajarin, CEO of technology consultancy Creative Strategies.
He added that Nvidia's pitch that the AI industry was going to be built on their chips in the long run would not change investor expectations for the short term. "Their narrative remains really unchanged," he said.
Huang announced new chips, including its next GPU chip Blackwell Ultra, which will be available in the second half of this year, and feature more memory than the current generation of its flagship chip Blackwell, meaning it can support larger AI models.
He said Nvidia's chips have two main purposes: helping AI systems respond smartly to a huge number of users, and giving those responses as fast as possible. Huang argued that Nvidia's chips are the only ones that can do both.
"If you take too long to answer a question, the customer is not going to come back. This is like web search," he said.
He also revealed details of a chip system called Vera Rubin, which will succeed Blackwell and feature faster speeds. It will be released in the second half of 2026.
Huang said Rubin chips will be followed by Feynman chips, arriving in 2028.
The new chip releases come as Blackwell is coming to market slower than expected after a design flaw caused manufacturing problems. The broader AI industry last year grappled with delays in which the prior methods of feeding expanding troves of data into ever-larger data centers full of Nvidia chips had started to show diminishing returns.
Nvidia said last month orders for Blackwell were "amazing."
Huang also introduced a powerful new personal computer called DGX Workstation based on Blackwell chips, saying it will be made by Dell, Lenovo and HP, among others. The device, which follows a smaller desktop machine introduced earlier this year, is a challenge to some of Apple's (AAPL.O), opens new tab top-end Macs.
"This is what a PC should look like," Huang said, holding up a motherboard for one of the devices.
He announced new software called Dynamo, which Nvidia released for free and is meant to speed up the process of reasoning. Huang also announced that automaker General Motors (GM.N), opens new tab has selected Nvidia to build its self-driving car fleet.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Nvidia CEO Huang says chipmaker well positioned for shift in AI
Stephen Nellis
March 18, 202511:24 PM EDT
Purchase Licensing Rights, opens new tab
SAN JOSE, California, March 18 (Reuters) - Nvidia (NVDA.O), opens new tab CEO Jensen Huang said on Tuesday the company was well placed to navigate a shift in the artificial intelligence industry, in which businesses are moving from training AI models to getting detailed answers from them.
Huang, speaking at the company's annual software developer conference in San Jose, California, defended the company's lead in selling costly AI chips to customers, which has recently been questioned by investors after China's DeepSeek made a competitive chatbot with allegedly fewer AI chips.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
But his presentation failed to reassure investors. Nvidia shares fell 3.4%. The chip index (.SOX), opens new tab closed down 1.6%.
"Almost the entire world got it wrong," Huang said on stage at the conference, dressed in his usual black leather jacket and jeans. He called the conference "the Super Bowl of AI."
"The amount of computation we need as a result of agentic AI, as a result of reasoning, is easily 100 times more than we thought we needed this time last year," he said, referring to autonomous AI agents that require little human intervention for routine tasks.
Nvidia's big moneymaking chips face pressure from technological change as AI markets shift from "training," AI models such as chatbots on huge troves of data to make them smart, to "inference," which is when the model uses its intelligence to produce answers for users.
Much of Nvidia's success stems from the decade the Santa Clara, California-based company spent building software tools to woo AI researchers and developers - but it was Nvidia's data center chips, which sell for tens of thousands of dollars each, that accounted for the bulk of its $130.5 billion in sales last year.
Its stock has more than quadrupled in value over the past three years as the company powered the rise of advanced AI systems such as ChatGPT, Claude and many others.
"I think the investor direction was that a lot of this news was priced in," said Ben Bajarin, CEO of technology consultancy Creative Strategies.
He added that Nvidia's pitch that the AI industry was going to be built on their chips in the long run would not change investor expectations for the short term. "Their narrative remains really unchanged," he said.
Huang announced new chips, including its next GPU chip Blackwell Ultra, which will be available in the second half of this year, and feature more memory than the current generation of its flagship chip Blackwell, meaning it can support larger AI models.
He said Nvidia's chips have two main purposes: helping AI systems respond smartly to a huge number of users, and giving those responses as fast as possible. Huang argued that Nvidia's chips are the only ones that can do both.
"If you take too long to answer a question, the customer is not going to come back. This is like web search," he said.
He also revealed details of a chip system called Vera Rubin, which will succeed Blackwell and feature faster speeds. It will be released in the second half of 2026.
Huang said Rubin chips will be followed by Feynman chips, arriving in 2028.
The new chip releases come as Blackwell is coming to market slower than expected after a design flaw caused manufacturing problems. The broader AI industry last year grappled with delays in which the prior methods of feeding expanding troves of data into ever-larger data centers full of Nvidia chips had started to show diminishing returns.
Nvidia said last month orders for Blackwell were "amazing."
Huang also introduced a powerful new personal computer called DGX Workstation based on Blackwell chips, saying it will be made by Dell, Lenovo and HP, among others. The device, which follows a smaller desktop machine introduced earlier this year, is a challenge to some of Apple's (AAPL.O), opens new tab top-end Macs.
"This is what a PC should look like," Huang said, holding up a motherboard for one of the devices.
He announced new software called Dynamo, which Nvidia released for free and is meant to speed up the process of reasoning. Huang also announced that automaker General Motors (GM.N), opens new tab has selected Nvidia to build its self-driving car fleet.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Unknown Author
March 18, 20255:17 AM EDT
Purchase Licensing Rights, opens new tab
BEIJING, March 18 (Reuters) - Tech giant Tencent (0700.HK), opens new tab on Tuesday unveiled a suite of new artificial intelligence tools capable of converting text and images into 3D visuals in the latest example of growing Chinese momentum in the field of generative AI.
Tencent released five open-source models based on its Hunyuan3D-2.0 technology, it announced in a statement, including so-called "turbo" versions it said can generate 3D visuals within 30 seconds while maintaining high precision and quality.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The launch comes as Chinese firms, notably led by AI startup DeepSeek, have begun challenging U.S. dominance in the field by offering comparable or superior performance at significantly lower costs.
Tencent's 3D initiative follows its February launch of Hunyuan Turbo S, a large language model that the company claims processes queries faster than DeepSeek's flagship R1 model.
Tencent first introduced its 3D AI models in November 2024, primarily targeting designers and game developers as it sought to position itself as an early mover in text-to-3D and image-to-3D generation capabilities.
The company launched Hunyuan3D-2.0 in January, claiming it outperforms leading industry models in text consistency, geometric accuracy, and visual quality benchmarks.
TikTok parent company ByteDance is also pushing into text-to-3D and image-to-3D field with its VeOmniverse model released in 2024. Unlike Tencent's offering, however, it remains proprietary.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Tencent expands AI push with open-source 3D generation tools
Unknown Author
March 18, 20255:17 AM EDT
Purchase Licensing Rights, opens new tab
BEIJING, March 18 (Reuters) - Tech giant Tencent (0700.HK), opens new tab on Tuesday unveiled a suite of new artificial intelligence tools capable of converting text and images into 3D visuals in the latest example of growing Chinese momentum in the field of generative AI.
Tencent released five open-source models based on its Hunyuan3D-2.0 technology, it announced in a statement, including so-called "turbo" versions it said can generate 3D visuals within 30 seconds while maintaining high precision and quality.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The launch comes as Chinese firms, notably led by AI startup DeepSeek, have begun challenging U.S. dominance in the field by offering comparable or superior performance at significantly lower costs.
Tencent's 3D initiative follows its February launch of Hunyuan Turbo S, a large language model that the company claims processes queries faster than DeepSeek's flagship R1 model.
Tencent first introduced its 3D AI models in November 2024, primarily targeting designers and game developers as it sought to position itself as an early mover in text-to-3D and image-to-3D generation capabilities.
The company launched Hunyuan3D-2.0 in January, claiming it outperforms leading industry models in text consistency, geometric accuracy, and visual quality benchmarks.
TikTok parent company ByteDance is also pushing into text-to-3D and image-to-3D field with its VeOmniverse model released in 2024. Unlike Tencent's offering, however, it remains proprietary.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Unknown Author
March 15, 202511:58 PM EDT
Purchase Licensing Rights, opens new tab
HONG KONG, March 16 (Reuters) - China's Baidu (9888.HK), opens new tab said on Sunday it has launched two new artificial intelligence models, including a new reasoning-focused model that it said rivalled DeepSeek's model, as it vies to stand out in a fiercely competitive AI race.
Chinese AI startup DeepSeek's roll-out of AI models which it says is on par with, or even better than, industry-leading models in the United States at a fraction of the cost, has roiled the industry and re-energised the global AI race.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
"ERNIE X1 delivers performance on par with DeepSeek R1 at only half the price," Baidu said of one of the new models. The X1 has "stronger understanding, planning, reflection, and evolution capabilities," Baidu said, adding that it is the first deep thinking model that uses tools autonomously.
Baidu said its latest foundation model ERNIE 4.5 has "excellent multimodal understanding ability. It has more advanced language ability, and its understanding, generation, logic, and memory abilities are comprehensively improved."
It also has "high EQ", and it is easy to understand network memes and satirical cartoons, Baidu said.
One of China's earliest tech giants to launch a ChatGPT-style chatbot, Baidu has struggled to gain widespread adoption for its Ernie large language model, despite claiming performance comparable to OpenAI's GPT-4, amid fierce competition.
Multimodal AI systems are capable of processing and integrating various types of data including text, video, images and audio, and can convert content across these formats.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
China's Baidu launches two new AI models as industry competition heats up
Unknown Author
March 15, 202511:58 PM EDT
Purchase Licensing Rights, opens new tab
HONG KONG, March 16 (Reuters) - China's Baidu (9888.HK), opens new tab said on Sunday it has launched two new artificial intelligence models, including a new reasoning-focused model that it said rivalled DeepSeek's model, as it vies to stand out in a fiercely competitive AI race.
Chinese AI startup DeepSeek's roll-out of AI models which it says is on par with, or even better than, industry-leading models in the United States at a fraction of the cost, has roiled the industry and re-energised the global AI race.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
"ERNIE X1 delivers performance on par with DeepSeek R1 at only half the price," Baidu said of one of the new models. The X1 has "stronger understanding, planning, reflection, and evolution capabilities," Baidu said, adding that it is the first deep thinking model that uses tools autonomously.
Baidu said its latest foundation model ERNIE 4.5 has "excellent multimodal understanding ability. It has more advanced language ability, and its understanding, generation, logic, and memory abilities are comprehensively improved."
It also has "high EQ", and it is easy to understand network memes and satirical cartoons, Baidu said.
One of China's earliest tech giants to launch a ChatGPT-style chatbot, Baidu has struggled to gain widespread adoption for its Ernie large language model, despite claiming performance comparable to OpenAI's GPT-4, amid fierce competition.
Multimodal AI systems are capable of processing and integrating various types of data including text, video, images and audio, and can convert content across these formats.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Stephen Nellis
March 18, 202512:02 PM EDT
Purchase Licensing Rights, opens new tab
SAN FRANCISCO, March 18 (Reuters) - Adobe on Tuesday said it is rolling out artificial intelligence "agents" that brands can use to help consumers navigate their websites.
Adobe is known for consumer apps such as Photoshop but also sells a suite of business-to-business software tools used in online marketing, which made up a quarter of Adobe's $21.5 billion in sales in its most recent fiscal year.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The new tools use AI technology to help brands target marketing messages to different users based on their online activity.
For example, content can be tailored to a young person's profile if they reached a website through a TikTok ad versus an older person's profile if they clicked on a result from a search query.
Amit Ahuja, senior vice president of Adobe's experience cloud unit, said users increasingly expect to interact with websites through chatbots as they do with apps such as ChatGPT.
Adobe's tools ensure that a business can offer that capability with a better awareness of the user.
For example, if a person lands on a travel booking site after clicking an ad on Instagram and asks about booking a trip, the site's chatbot can check both inventory as well as what kind of destination was shown in the ad, providing a better idea of what to suggest to the user.
Ahuja said AI functions help websites co-ordinate both functions, which were previously managed in completely different ways "with different guardrails".
Adobe is also releasing other new tools for digital marketers. One tool, for example, allows marketing professionals to tell AI agents their goals for making website changes to increase digital sales. The agent in response can recommend ways to get it done and then make the changes.
Ahuja said Adobe's aim is to speed up "what would have taken months in the past, where I'm waiting for a coding team or somebody to go fix it."
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Adobe rolls out AI agents for online marketing tools
Stephen Nellis
March 18, 202512:02 PM EDT
Purchase Licensing Rights, opens new tab
SAN FRANCISCO, March 18 (Reuters) - Adobe on Tuesday said it is rolling out artificial intelligence "agents" that brands can use to help consumers navigate their websites.
Adobe is known for consumer apps such as Photoshop but also sells a suite of business-to-business software tools used in online marketing, which made up a quarter of Adobe's $21.5 billion in sales in its most recent fiscal year.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The new tools use AI technology to help brands target marketing messages to different users based on their online activity.
For example, content can be tailored to a young person's profile if they reached a website through a TikTok ad versus an older person's profile if they clicked on a result from a search query.
Amit Ahuja, senior vice president of Adobe's experience cloud unit, said users increasingly expect to interact with websites through chatbots as they do with apps such as ChatGPT.
Adobe's tools ensure that a business can offer that capability with a better awareness of the user.
For example, if a person lands on a travel booking site after clicking an ad on Instagram and asks about booking a trip, the site's chatbot can check both inventory as well as what kind of destination was shown in the ad, providing a better idea of what to suggest to the user.
Ahuja said AI functions help websites co-ordinate both functions, which were previously managed in completely different ways "with different guardrails".
Adobe is also releasing other new tools for digital marketers. One tool, for example, allows marketing professionals to tell AI agents their goals for making website changes to increase digital sales. The agent in response can recommend ways to get it done and then make the changes.
Ahuja said Adobe's aim is to speed up "what would have taken months in the past, where I'm waiting for a coding team or somebody to go fix it."
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Tim McLaughlin
March 19, 20257:01 AM EDT
Purchase Licensing Rights, opens new tab
BOSTON, March 19 (Reuters) - Data Center Alley, a 30-square-mile stretch outside Washington D.C. and home to more than 200 data centers, consumes roughly the same electricity as Boston. So power company officials were alarmed when a big chunk of those centers - 60 of them - suddenly dropped off the grid one day last summer and switched to on-site generators.
The mass reaction was triggered by a standard safety mechanism across the data center industry, intended to protect computer chips and electronic equipment from damage caused by voltage fluctuations. But it caused a huge surge in excess electricity, according to federal regulators and utility executives.
The Technology Roundup newsletter brings the latest news and trends straight to your inbox. Sign up here.
The magnitude of the imbalance forced grid operator PJM and local utility Dominion Energy (D.N), opens new tab to scale back output from power plants to protect grid infrastructure and avoid a worst-case scenario of cascading power outages across the region.
The near-miss - reported here in detail for the first time - forced federal regulators to recognize a new vulnerability of America’s electrical grid: unannounced disconnections by data centers.
"As these data centers get bigger and consume more energy, the grid is not designed to withstand the loss of 1,500-megawatt data centers," John Moura, Director of Reliability Assessment and System Analysis for NERC, told Reuters in an interview. "At some level it becomes too large to withstand unless more grid resources are added."
Historically, grid operators have planned for large power plants tripping offline. But the rapid expansion of data centers processing the vast amounts of information used for AI and crypto mining is forcing grid operators to plan for new contingencies and complicating the already difficult task of balancing the country’s supply and demand of electricity.
"What it tells us is that the behavior of data centers has the potential to cause cascading power outages for an entire region," said Alison Silverstein, a former senior adviser to the chairman of the U.S. Federal Energy Regulatory Commission.
The event last July 10 occurred near the D.C. suburb of Fairfax, Virginia, an area known as Data Center Alley for its concentration of facilities serving Microsoft, Google and Amazon. About 70% of the world's internet traffic flows through the area.
A month after the incident, the North American Electric Reliability Corporation (NERC), the federal regulator for grid reliability, founded a taskforce to study en masse disconnections by data centers and crypto miners.
For this story, Reuters examined thousands of pages of regulatory documents and interviewed about a dozen industry executives to determine the origins of the fault - a failed surge protector on Dominion's Ox-Possum 230-kilovolt line near Fairfax, Virginia – and its spread across the area.
NERC reviewed the incident in a report in January but did not disclose the exact location of the fault, the number of data centers involved, or how PJM and Dominion worked to rebalance the grid’s supply and demand of electricity.
The number of near-miss events like the one in Data Center Alley has grown rapidly over the last five years as more data centers come online.
The amount of power used by data centers has tripled over the past decade and could triple again by 2028, according to a report produced by the Lawrence Berkeley National Laboratory for the Department of Energy in December.
A Reuters review of disclosure filings by the Electric Reliability Council of Texas (ERCOT), the state’s main grid operator, identified more than 30 near-miss incidents since 2020, triggered by big energy users like data centers and crypto miners switching offline.
In December 2022, a failed transformer at a substation in west Texas caused nearly 400 crypto miners, data centers and oil and gas production facilities to unplug without warning.
The mass exodus produced an oversupply of nearly 1,700 megawatts of electricity – equivalent to about about 5% of the grid's total demand - and forced 112 megawatts of power generation to shut down, according to ERCOT.
The risk of power outages will only grow as new data centers come online, the NERC forecast in a December report. Nearly all of the United States will face higher risks of energy shortfalls over the next 5 to 10 years, the report said.
The regulator urged utilities to consider updating federal reliability standards for data centers and crypto miners.
Many data centers are engineered by their operators to switch to local generators at the smallest hint of a problem on the grid to minimize the risk of an interruption to services like Google search or crypto mining, according to NERC.
Some grid operators have proposed requiring data centers to “ride through” routine voltage dips without disconnecting. But data center operators are opposed because of the risk of damaging electronic equipment and cooling systems.
ERCOT last year withdrew a proposal that would have imposed ride-through restrictions on data centers and crypto miners after facing pushback from an industry group, the Data Center Coalition.
The group, whose members include Amazon, Google, and Meta, cited costs and the risk of damaging computer chips and cooling systems exposed to fluctuating voltage levels.
"Data center hardware and power supplies, similar to other electronics, are very sensitive to power supply stability," the coalition said in January 2024 comments filed with ERCOT.
"Deviating from this range will deteriorate the optimal performance, reduce longevity, or damage the components beyond repair."
The coalition declined to comment for this story. Amazon, Google and Meta did not return messages seeking comment. ERCOT did not return messages seeking comment.
There is "high potential" for the magnitude of these disconnection events to grow as larger operations plug into the Texas grid, ERCOT operations engineer Patrick Gravois said in a December presentation to NERC’s Large Load Task Force.
Gravois said the grid operator is still working to determine exactly what prompts big users of electricity to unplug from the grid, so that it can avoid surprises.
Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School, said regulators could require data centers to ride through voltage dips - but that could risk Big Tech decamping for states with more relaxed rules.
Jim Simonelli, chief technology officer for Schneider Electric’s secure power division, said utilities and the data center industry have a lot of lessons to be learned from what happened outside Washington DC this past July.
"One thing that doesn’t exist yet for the data center industry is how to be grid-friendly,” Simonelli said.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
Big Tech’s data center boom poses new risk to US grid operators
Tim McLaughlin
March 19, 20257:01 AM EDT
Purchase Licensing Rights, opens new tab
BOSTON, March 19 (Reuters) - Data Center Alley, a 30-square-mile stretch outside Washington D.C. and home to more than 200 data centers, consumes roughly the same electricity as Boston. So power company officials were alarmed when a big chunk of those centers - 60 of them - suddenly dropped off the grid one day last summer and switched to on-site generators.
The mass reaction was triggered by a standard safety mechanism across the data center industry, intended to protect computer chips and electronic equipment from damage caused by voltage fluctuations. But it caused a huge surge in excess electricity, according to federal regulators and utility executives.
The Technology Roundup newsletter brings the latest news and trends straight to your inbox. Sign up here.
The magnitude of the imbalance forced grid operator PJM and local utility Dominion Energy (D.N), opens new tab to scale back output from power plants to protect grid infrastructure and avoid a worst-case scenario of cascading power outages across the region.
The near-miss - reported here in detail for the first time - forced federal regulators to recognize a new vulnerability of America’s electrical grid: unannounced disconnections by data centers.
"As these data centers get bigger and consume more energy, the grid is not designed to withstand the loss of 1,500-megawatt data centers," John Moura, Director of Reliability Assessment and System Analysis for NERC, told Reuters in an interview. "At some level it becomes too large to withstand unless more grid resources are added."
Historically, grid operators have planned for large power plants tripping offline. But the rapid expansion of data centers processing the vast amounts of information used for AI and crypto mining is forcing grid operators to plan for new contingencies and complicating the already difficult task of balancing the country’s supply and demand of electricity.
"What it tells us is that the behavior of data centers has the potential to cause cascading power outages for an entire region," said Alison Silverstein, a former senior adviser to the chairman of the U.S. Federal Energy Regulatory Commission.
The event last July 10 occurred near the D.C. suburb of Fairfax, Virginia, an area known as Data Center Alley for its concentration of facilities serving Microsoft, Google and Amazon. About 70% of the world's internet traffic flows through the area.
A month after the incident, the North American Electric Reliability Corporation (NERC), the federal regulator for grid reliability, founded a taskforce to study en masse disconnections by data centers and crypto miners.
For this story, Reuters examined thousands of pages of regulatory documents and interviewed about a dozen industry executives to determine the origins of the fault - a failed surge protector on Dominion's Ox-Possum 230-kilovolt line near Fairfax, Virginia – and its spread across the area.
NERC reviewed the incident in a report in January but did not disclose the exact location of the fault, the number of data centers involved, or how PJM and Dominion worked to rebalance the grid’s supply and demand of electricity.
The number of near-miss events like the one in Data Center Alley has grown rapidly over the last five years as more data centers come online.
The amount of power used by data centers has tripled over the past decade and could triple again by 2028, according to a report produced by the Lawrence Berkeley National Laboratory for the Department of Energy in December.
A Reuters review of disclosure filings by the Electric Reliability Council of Texas (ERCOT), the state’s main grid operator, identified more than 30 near-miss incidents since 2020, triggered by big energy users like data centers and crypto miners switching offline.
In December 2022, a failed transformer at a substation in west Texas caused nearly 400 crypto miners, data centers and oil and gas production facilities to unplug without warning.
The mass exodus produced an oversupply of nearly 1,700 megawatts of electricity – equivalent to about about 5% of the grid's total demand - and forced 112 megawatts of power generation to shut down, according to ERCOT.
The risk of power outages will only grow as new data centers come online, the NERC forecast in a December report. Nearly all of the United States will face higher risks of energy shortfalls over the next 5 to 10 years, the report said.
The regulator urged utilities to consider updating federal reliability standards for data centers and crypto miners.
Many data centers are engineered by their operators to switch to local generators at the smallest hint of a problem on the grid to minimize the risk of an interruption to services like Google search or crypto mining, according to NERC.
Some grid operators have proposed requiring data centers to “ride through” routine voltage dips without disconnecting. But data center operators are opposed because of the risk of damaging electronic equipment and cooling systems.
ERCOT last year withdrew a proposal that would have imposed ride-through restrictions on data centers and crypto miners after facing pushback from an industry group, the Data Center Coalition.
The group, whose members include Amazon, Google, and Meta, cited costs and the risk of damaging computer chips and cooling systems exposed to fluctuating voltage levels.
"Data center hardware and power supplies, similar to other electronics, are very sensitive to power supply stability," the coalition said in January 2024 comments filed with ERCOT.
"Deviating from this range will deteriorate the optimal performance, reduce longevity, or damage the components beyond repair."
The coalition declined to comment for this story. Amazon, Google and Meta did not return messages seeking comment. ERCOT did not return messages seeking comment.
There is "high potential" for the magnitude of these disconnection events to grow as larger operations plug into the Texas grid, ERCOT operations engineer Patrick Gravois said in a December presentation to NERC’s Large Load Task Force.
Gravois said the grid operator is still working to determine exactly what prompts big users of electricity to unplug from the grid, so that it can avoid surprises.
Ari Peskoe, director of the Electricity Law Initiative at Harvard Law School, said regulators could require data centers to ride through voltage dips - but that could risk Big Tech decamping for states with more relaxed rules.
Jim Simonelli, chief technology officer for Schneider Electric’s secure power division, said utilities and the data center industry have a lot of lessons to be learned from what happened outside Washington DC this past July.
"One thing that doesn’t exist yet for the data center industry is how to be grid-friendly,” Simonelli said.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
reuters.com
Unknown Author
March 11, 20252:42 PM EDT
Purchase Licensing Rights, opens new tab
March 11 (Reuters) - OpenAI launched new tools for developers on Tuesday that will help them build advanced AI agents, using a few application programming interfaces (APIs), amid growing competition from Chinese AI startups.
AI agents are designed to independently execute complex real-world tasks without direct human intervention, while API is a strings of code which enables standardized communication, data exchange and functionality between software components.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The new tool, called the Responses API, is available to all developers at no additional cost. It replaces OpenAI's Assistants API, which is set to be phased out by the second half of 2026.
The development comes close on the heels of release of latest AI models by Chinese startups which claim to be on par or better than industry-leading models in the United States at a fraction of the cost.
Chinese startup Monica has grabbed attention over the past few days after its launched its autonomous AI agent Manus, weeks after DeepSeek was showered with praise by Silicon Valley executives and U.S. tech company engineers.
Monica, which claims that its Manus AI outperforms OpenAI's DeepResearch agent, said on Tuesday that it was partnering with the team behind Alibaba's Qwen AI models.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab
OpenAI launches new developer tools as Chinese AI startups gain ground
Unknown Author
March 11, 20252:42 PM EDT
Purchase Licensing Rights, opens new tab
March 11 (Reuters) - OpenAI launched new tools for developers on Tuesday that will help them build advanced AI agents, using a few application programming interfaces (APIs), amid growing competition from Chinese AI startups.
AI agents are designed to independently execute complex real-world tasks without direct human intervention, while API is a strings of code which enables standardized communication, data exchange and functionality between software components.
The Reuters Daily Briefing newsletter provides all the news you need to start your day. Sign up here.
The new tool, called the Responses API, is available to all developers at no additional cost. It replaces OpenAI's Assistants API, which is set to be phased out by the second half of 2026.
The development comes close on the heels of release of latest AI models by Chinese startups which claim to be on par or better than industry-leading models in the United States at a fraction of the cost.
Chinese startup Monica has grabbed attention over the past few days after its launched its autonomous AI agent Manus, weeks after DeepSeek was showered with praise by Silicon Valley executives and U.S. tech company engineers.
Monica, which claims that its Manus AI outperforms OpenAI's DeepResearch agent, said on Tuesday that it was partnering with the team behind Alibaba's Qwen AI models.
Our Standards: The Thomson Reuters Trust Principles., opens new tab, opens new tab