Latest / In The Loop / The Crazy Story Of How OpenAI Lost Everything In One Week

Transcript

The Crazy Story Of How OpenAI Lost Everything In One Week · In The Loop

• 0:00
  This is the story of how in just ten days, Open AI went from
• 0:03
  being close to its biggest acquisition in history and being
• 0:06
  able to truly compete in the AI coding arena, to being beaten by
• 0:11
  Elon Musk and Grock's new large language model, losing most of
• 0:15
  its AI researchers and watching the company it tried to acquire
• 0:20
  get bought out by Google instead in what has been coined the
• 0:24
  biggest acquisition in history. To see how we got here, let's go
• 0:27
  back just a few weeks. This is In the Loop with Jack
• 0:30
  Horton. I hope you enjoy the show.
• 0:43
  So this started in May 2025 when Sam Altman and Open AI made what
• 0:48
  appeared to be potentially quite a clever and ambitious
• 0:51
  acquisition of Winsurf AI, which is a coding platform.
• 0:56
  This would have been the largest Open AI acquisition to date,
• 0:58
  about $3 billion. So a huge sum of money.
• 1:01
  Many people called it overpriced, many people called
• 1:04
  it clever. But now we never really get to
• 1:07
  see whether that bet paid off. Now, it's really important not
• 1:10
  to forget how important the AI coding industry is to these big
• 1:14
  players because it's one of the most successful areas of the
• 1:18
  world where AI has been adopted and people are paying huge sums
• 1:21
  of money for it. And competition over the last 12
• 1:24
  months has become crazy. So Cursor just raised about 900
• 1:28
  million dollars, $10 billion valuation.
• 1:31
  And Anthropic has been positioning Cloud four as
• 1:35
  basically the best coding tool in the world.
• 1:36
  And they have literally been adopted by every major coding
• 1:39
  platform like livable Cursor Replit as the base underlying
• 1:44
  large language model. And so for Open AI, the
• 1:47
  acquisition made quite a lot of sense.
• 1:49
  First, it would obviously bring agent based coding directly into
• 1:53
  their platform by integrating much of the Windsurf technology
• 1:57
  into the Chatcher BT technology. You know apparently Windsurf had
• 2:01
  won a lot of major enterprise customers, you know Dell, JP
• 2:04
  Morgan and they had about $40 million in ARR, so annualized
• 2:09
  recurring revenue and the company had raised over $200
• 2:12
  million and reached about a $1.25 billion valuation just
• 2:16
  2020 follow them. So again, it made some sense to
• 2:19
  acquire and build this into their product.
• 2:22
  Second, it would close what has been described by many as a big
• 2:26
  feature gap from the likes of Open AI.
• 2:28
  Because startups like Windsurf and rivals like Cursor, we're
• 2:32
  offering developers much better experiences on their product.
• 2:35
  Again, don't forget how important it is to have the
• 2:37
  developer community using your technology in large language
• 2:41
  models to build AI applications. Third, it was what could be
• 2:45
  described as quite a defensive play as well because it
• 2:47
  prevented competitors like Google or others acquiring the
• 2:50
  technology and trying to do the exact same.
• 2:53
  So this is when things went South
• 3:05
  under Open AI's existing partnership with Microsoft.
• 3:09
  As you may know, Microsoft owned about 49% of the company and are
• 3:13
  one of its main funders. Well, under that agreement, all
• 3:17
  the rights to Open AI's technology goes to Microsoft.
• 3:21
  So Microsoft is allowed to use any of the open AI technology
• 3:24
  developed, which would have also meant any acquisitions they
• 3:27
  made. Meaning Microsoft would
• 3:29
  automatically get access to all of Winsurf's proprietary
• 3:32
  technology. Something that the founders of
• 3:35
  Winsurf, as you can imagine, were vehemently against.
• 3:39
  Because from Windsor's perspective, Microsoft was an
• 3:43
  existential threat. Their product, GitHub copilot,
• 3:46
  was quite literally one of their number one competitors.
• 3:48
  So being acquired by a big company without any IP
• 3:51
  protection rights was a big no no for the founders.
• 3:54
  And from Microsoft's perspective, how could they let
• 3:57
  one of their partners buy a company that would take away its
• 4:02
  main coding product or could take away it's main main coding
• 4:05
  product? Because GitHub Copilot is $10 a
• 4:08
  user and they've got millions of users.
• 4:09
  Now, according to many reports, Open AI did attempt to negotiate
• 4:14
  a carve out that would basically exempt Windsurf's IP from any of
• 4:18
  these agreements, but Microsoft completely refused.
• 4:21
  The idea kind of came at a bad time for Open AI because
• 4:25
  Microsoft and Open AI are in the midst of a new partnership
• 4:29
  agreement on what the future of their their partnership looks
• 4:32
  like. And any budget from Microsoft in
• 4:35
  those negotiations could mean further budgets in other deal
• 4:38
  terms. So again, Open AI and Microsoft
• 4:42
  weren't budging because from Microsoft's perspective, the
• 4:44
  stance made sense. They don't want to give away
• 4:46
  potential. Now a lot of this has got quite
• 4:49
  tense as you can imagine because according to Axios, Open AI even
• 4:53
  were considering what's described as a nuclear option,
• 4:56
  which is accusing Microsoft of antitrust violations as part of
• 5:00
  their broader agreement that was happening.
• 5:03
  As I said at the same time, again, the antitrust part being
• 5:07
  you are threatening our business success in favor of yours.
• 5:11
  So the windsurf situation got a lot worse in June because
• 5:14
  Anthropic then decided to cut off access to Windsurf's usage
• 5:19
  of its coding model. Again, Anthropic who is part
• 5:23
  owned by the likes of Google and Amazon.
• 5:25
  And again, don't forget that the cloud models were by a mile
• 5:28
  developers favorite model to use all because Anthropic said if
• 5:31
  they were going to become part of the open AI ecosystem, then
• 5:35
  all access to their models were turned off.
• 5:37
  So here's the next part in that big story, which is I guess the
• 5:40
  entrance of Google, because open AI is exclusive negotiation
• 5:43
  window ended in July 2025 and the dealer officially collapsed.
• 5:48
  And within days of that happening, Google DeepMind
• 5:51
  basically has executed what's been described as one of the
• 5:54
  biggest acquisitions in history by essentially, instead of
• 5:58
  purchasing Windsurf, acquiring their technology through a
• 6:01
  licensing agreement and hiring all of Windsurfers team and
• 6:05
  bringing them into the fold of DeepMind.
• 6:07
  So Google's AI Lab. And Google announced that it
• 6:10
  hired the CEO of Windsurf as well as several top engineers to
• 6:15
  come directly into the Deepmine team and lead the AI coding
• 6:19
  division within the Gemini so language model, Gemini's
• 6:23
  department. Now, this is super clever by
• 6:25
  Google. Rather than acquiring it for say
• 6:27
  $3 billion like Open Eye had negotiated, Google paid about
• 6:31
  $2.4 billion for what's called a non exclusive license to
• 6:35
  Windsurf's technology, which basically means that other
• 6:38
  people can use the technology, but they have a license to it.
• 6:42
  And this covered a multi year compensation for all the staff
• 6:45
  that they hired. And the reason this is so clever
• 6:48
  is because they're not taking an equity stake in the platform
• 6:53
  Windsurf itself. They essentially enabled
• 6:56
  Windsurf to remain independent, keep innovating, but from
• 7:01
  Google's perspective, avoided any regulatory scrutiny because
• 7:05
  every time a big tech giant pays $3 billion for another product,
• 7:09
  they typically then get investigated.
• 7:11
  Now, if this wasn't bad enough for the likes of Open Eye, who'd
• 7:13
  have made a huge bet like spending $3 billion on a big
• 7:16
  company isn't just this small decision that happens in a few
• 7:18
  weeks. It's a very long process, and to
• 7:22
  make matters worse, over the last 10 days Open AI had had a
• 7:25
  series of unfortunate events happen in the world.
• 7:38
  Because if that hadn't really annoyed Sam Altman enough, Elon
• 7:43
  Musk then launched the XA is latest Grok model.
• 7:48
  Now, you may have seen this all over the news.
• 7:49
  I nearly did this entire episode on that single model release,
• 7:52
  but I thought it'd be interesting to wrap it into this
• 7:55
  discussion too. Because Elon Musk's XAI, just
• 7:58
  two days before this announcement of the windsurf
• 8:00
  situation, launched what's called Grok 4, which is a model
• 8:03
  that has surpassed even GPTS to open a eyes.
• 8:06
  Best models across a number of the most respected benchmarks
• 8:11
  and even independent benchmarks. And now arguably and
• 8:14
  technically, it is the best large language model in the
• 8:16
  world, which to me is just hilarious and it's madness.
• 8:20
  You know it posted best in class scores across coding reasoning
• 8:25
  math almost everything apart from things like speed and cost
• 8:28
  in which it is more expensive and slower than in terms of its
• 8:31
  responses than other models. An independent evaluators
• 8:34
  confirm that XA is claims were justified because the
• 8:38
  benchmarking project artificial analysis reported that its
• 8:42
  overall model performance surpassed both Gemini 2.5 and
• 8:46
  all of Open A is best offerings. And obviously, as you can
• 8:50
  imagine, Elon Musk wasn't modest about the achievements as he
• 8:53
  said it's the smartest AI in the world.
• 8:55
  And basically he's the best and the greatest of course.
• 8:58
  But he actually credited Grock's performance with what's called a
• 9:02
  tool integrated reasoning. So the ability to use tools and
• 9:05
  do reasoning, much like the open AI models, to be honest, but
• 9:09
  actually one of the biggest training runs.
• 9:11
  So 2.4 trillion parameter scale training data and training runs
• 9:17
  for for that incredible frontier performance and the fact that
• 9:20
  open AI don't really have a response.
• 9:21
  There's a lot of noise and complaints about the fact that
• 9:24
  GPT 5 still isn't out and there's not much news about it.
• 9:27
  And apparently Open AI haven't spent 18 months trying to get
• 9:31
  this right and hundreds of $1,000,000 on at least two
• 9:34
  training runs for GPT 5 to only find the results underwhelming.
• 9:38
  Again, it's just more bad news for open AI.
• 9:52
  Now obviously these two things are very annoying for the likes
• 9:54
  of Sam Altman and Open AI, but to make matters worse, and yes
• 9:58
  it does get worse, in just the span of 20-30 days, Meta.
• 10:03
  So Zuckerberg has poached over 20 Senior Open AI researchers in
• 10:08
  what can only be described as probably the most aggressive
• 10:12
  talent heist ever. So at the end of June, on June
• 10:15
  26th, the Wall Street Journal reported that Meta but snapped
• 10:19
  up open AI's vision researchers. So those researchers that are a
• 10:23
  key part of making language models understand things in the
• 10:26
  world. The people behind much of the
• 10:29
  voice work, also the image pipeline and even the O3 model
• 10:35
  creators as well. And Zuckerberg has called this
• 10:37
  the Super intelligence lab. Now, Sam Altman told Reuters
• 10:40
  that Mehta was offering a signing up package of up to $100
• 10:44
  million for attracting these Meta researchers.
• 10:46
  And yes, you heard that right, $100 million signing on bonus.
• 10:50
  And arguably the real draw outside of the $100 million,
• 10:53
  which is a big draw, is the promise of basically limited
• 10:56
  resources, access to the very best training hardware for AI
• 11:00
  models, plus unlimited freedom from any corporate constraints.
• 11:05
  This is basically a team that can do whatever they want.
• 11:08
  They've don't have to worry about consumer products and
• 11:11
  revenue, they've just been tasked with building a model
• 11:14
  that beats everything in the world.
• 11:16
  Open source weights and move really fast.
• 11:21
  So again, no consumer road map to build for, no Microsoft
• 11:25
  vetoing power that drives the researchers up the wall, No
• 11:29
  budget committees to convince. Just pure research backed by
• 11:33
  unlimited compute, which to researchers trying to create AGI
• 11:38
  and incredible things must be incredibly attractive.
• 11:50
  So what does this all really mean?
• 11:52
  Well, the talent exodus probably couldn't have come at a worst
• 11:55
  time for open AI. They've just lost a load of
• 11:57
  their best researchers who were the architects of Open A is most
• 12:01
  important products over the last few years.
• 12:05
  The narrative damage to Open AI after Google snatched up
• 12:09
  Windsurf is incredibly negative and just a failure to acquire
• 12:14
  this company because of Microsoft only probably shines a
• 12:16
  light on future conflicts and tensions with the Microsoft Open
• 12:20
  AI partnership. And for the first time ever
• 12:23
  since Chatt GPT launched, Open AI looks more vulnerable not
• 12:27
  just to, I guess specific competitive threats, but to a
• 12:32
  fundamental erosion of its competitive advantages.
• 12:35
  Having the best models, the best researchers, the best people.
• 12:39
  It still however has the most users, but again, things like
• 12:42
  that can change really quick. Anyway, that's my deep dive into
• 12:47
  the Open AI saga and I hope that was interesting.
• 12:49
  Otherwise, I'll see you next week.