# Elon Raises $40B for GPUs, Anthropic's Enterprise Strategy, Microsoft Lays offs 60K

An episode of [AI Chat: AI News on Claude, ChatGPT, Gemini, Grok, Perplexity](https://invisiblenetwork.ai/podcast/1669990482), 27 Oct 2026, 15 min.

Audio: https://content.rss.com/episodes/365073/3268017/ai-chat-podcast/2026_10_26_22_45_22_93be527b-bfd2-46d3-bbf9-4612891eef2c.mp3

Got something to say to this podcast? Send a voicemail, and you might hear yourself on the next episode! In this episode, we discuss Elon Musk's recent $40 billion fundraising effort aimed at acquiring GPUs. We also cover Microsoft's planned layoff of 60,000 employees and Anthropics' strategy to retain enterprise customers. Show Links • Get the top 80+ AI Models for $8.99 at AI Box: https://aibox.ai • How I Grow and Scale My Business with AI: https://www.skool.com/aihustle • Get the AI Chat Daily Newsletter: https://www.aichatdaily.com/newsletter

## Transcript

[0:00] Welcome to the podcast. Today on the show, Anthropic
[0:02] has announced their IPO. And in their filing,
[0:06] I want to break down a couple interesting things
[0:07] that are showing their strategy for keeping their
[0:10] enterprise customers hooked. There's an interesting
[0:12] article in TechCrunch on that. At the same time,
[0:15] we have Elon Musk that says SpaceX has just closed
[0:18] $40 billion in debt financing to buy AI GPUs.
[0:22] We have Microsoft that is offering a voluntary
[0:26] buyout to about 60 ,000 employees, which is a
[0:30] quarter of its entire staff. And then we have
[0:33] Google DeepMind that has a model that correctly
[0:36] called when Hurricane Isaiah was going to make
[0:39] landfall in Florida. We have SoftBank's Masoyo.
[0:45] Masio Shisan, who's planning a $50 billion AI
[0:49] and data center push in Malaysia. And then we
[0:52] have Kohir that says Alibaba's Qen 3 .8 Flash
[0:57] Next tried to upload a file to a Beijing cloud
[1:00] bucket, which I've interviewed some of the co
[1:02] -founders of Kohir. And so based off of what
[1:05] they're doing and with ai sovereignty this is
[1:08] a really big deal and of course for security
[1:10] and all of that before we get into it if you
[1:12] want to try any of these models that i just mentioned
[1:14] from google to microsoft to OpenAI, to Anthropic.
[1:18] Ko here, I would love for you to check out my
[1:20] new, my startup, which is AIbox .ai. We have
[1:23] a brand new feature we just released that allows
[1:26] you to generate PowerPoints and documents and
[1:29] all sorts of rich assets inside of our playground,
[1:32] which you can chat with over a hundred different
[1:34] AI models, everything from image, audio, video,
[1:37] text, music. For $8 .99, you can talk to all
[1:40] of those different AI models in one thread. You
[1:42] can make a group of your favorites. So you might
[1:45] do something like ChatGPT, Claude, Gemini, and
[1:47] Grok. Ask one question. All four of them answer
[1:50] it at the same time. You can pick which one you
[1:51] like the best. If you're making some business
[1:53] document, you can pull in the response that you
[1:55] like best. You could also do that with image
[1:56] generators. Sometimes I generate an image with
[1:58] one model and I'm like, man, ChatGPT did a bad
[2:00] job on that. I'll go do something like Ideogram
[2:03] or some of these other. innovative AI models.
[2:05] But if you make a group of four of those, you
[2:07] tell it to generate an image, you just pick the
[2:09] one that you like the best. Sometimes when you're
[2:11] in chat, you're like, no, generate it again,
[2:13] but make it more realistic and generate it again,
[2:14] but make it more like this. If you don't want
[2:16] to have to do that over and over again, just
[2:17] pick four different models. And one of those
[2:19] models will oftentimes generate that exact spec
[2:22] for an image the way that you want it the first
[2:24] time. So this is one of the one of the strategies
[2:26] I've been deploying. If you want to check it
[2:28] out, it is AI box .ai. I'd love for you to go
[2:31] and try it out. The first thing I want to cover
[2:33] here is the story with SpaceX. So this actually
[2:36] came out of a tweet that I saw earlier today
[2:38] from Elon Musk. But basically, he said that SpaceX
[2:42] just closed $40 billion in debt financing, and
[2:44] they're doing this to buy GPUs for AI data centers.
[2:46] Now, this isn't anything crazy. Everyone's been
[2:49] doing this for a long time. All of these neoclouds
[2:51] have been doing this for a long time, getting
[2:52] debt, buying GPUs, building data centers. But
[2:55] he said that the debt is callable and it's going
[2:58] to fund the GPU deliveries that are due in Q1
[3:01] and Q2 of next year. So he said this is going
[3:03] to have no impact on what they're currently doing,
[3:05] their current quarter. This is just debt that
[3:07] they're getting for their next delivery. SpaceX
[3:10] guided up to $35 billion to $40 billion in Q4
[3:14] CapEx and about 200 ,000 GB300 chips coming online
[3:18] each month. And they're scaling very rapidly
[3:21] right now. Their Q3 CapEx was $19. $16 .6 billion.
[3:25] That included $16 .4 billion for SpaceX's super
[3:28] intelligence business. Elon Musk said that SpaceX
[3:32] will not buy any neoclouds. He predicted there's
[3:35] going to be a big consolidation among a lot of...
[3:37] poorly capitalized GPU rental providers, which
[3:40] if you know basically how this works, there's
[3:42] tons of these companies that will go buy GPUs.
[3:44] They're called quote unquote neoclouds. And then
[3:46] once they have acquired them, they're going to
[3:49] go and just rent it out to different startups.
[3:52] And you're able to go and essentially, you know,
[3:55] pay them for some bit of compute. Now, Elon had
[3:59] a whole thread over on Twitter. where he's talking
[4:02] about what's going on with that compute. He said,
[4:05] I think we're going to see a lot of stranded
[4:07] compute in partially complete data centers. So
[4:11] these data centers that are not fully built out,
[4:13] there's a lot of kind of pushback on a lot of
[4:15] data centers and a lot of legal stuff going on
[4:17] there. And Chamath Palihapitiya from the All
[4:19] In podcast, who invests in this, responded with
[4:22] a really interesting take. He said, as someone
[4:24] with a portfolio of energized land assets, here
[4:26] is what I am seeing. Neoclouds are disorganized
[4:29] and poorly capitalized. Some are definitely better
[4:31] than others, however. Hyperscales are irretractable
[4:35] bureaucratic organizations that take three months
[4:37] to do an NDA and want an 18 month escrow period.
[4:41] We build our portfolio by competing against these
[4:43] terms because of our speed and simplicity. The
[4:45] more future assets orgs like ours win, the bigger
[4:49] the problem for the hyperscalers downstream.
[4:51] He said there are delays in the existing pipeline
[4:53] that will catch people off guard in 2027. Hyperscalers
[4:57] will obviously react by focusing on their own
[4:59] internal compute needs over any third party.
[5:01] So you can see people like Amazon and Google
[5:04] that are. hyperscalers they like google needs
[5:06] they have google cloud they they they're selling
[5:10] their compute to startups but if google really
[5:12] was in a pinch and they needed all of the compute
[5:15] they would keep it for themselves they wouldn't
[5:16] give it to any of these startups right they're
[5:18] putting themselves obviously ahead of all of
[5:20] their competitors he said if the third party
[5:22] is a model company they will tell them to go
[5:25] through their model router where they will also
[5:27] take a non -trivial vig them telling anthropic
[5:30] that they have no compute for them rumored in
[5:33] anthropics s1 if it's true is a huge tell that
[5:35] this is already happening all of this is to say
[5:37] that it doesn't look like there will be a lot
[5:39] of partially surrounded compute and compute sitting
[5:42] in inventory in 2027 that is currently forecasted
[5:44] this is bought and paid for but effectively under
[5:47] earning below total cost of capital or collecting
[5:50] dust we see this as true for everyone which Everyone,
[5:54] model companies, hyperscalers, neoclouds, except
[5:56] one. You can guess which one. And I think he's
[5:58] alluding to SpaceX. He said this dislocation
[6:01] will benefit two kinds of organizations, the
[6:03] non -bureaucratic big scalers who can build industrial
[6:06] scale quickly and the nimble, well -capitalized
[6:09] players feeding them. 2027 will be a turbulent
[6:11] year because it will clear out a bunch of wrong,
[6:14] bad bets. It will also. be a consolidation year
[6:16] for a few. Elon Musk said pretty much as of the
[6:19] last few weeks, debt markets have almost entirely
[6:22] shut off access to money for any company that
[6:25] doesn't have. substantial revenue independent
[6:27] of AI data centers. SpaceX, this is when he mentioned
[6:30] that they closed their $40 billion of debt for
[6:32] GPUs. He says, but we're investment grade and
[6:35] we have Starlink and launch revenue. That means
[6:37] that 200 plus Neo clouds are going to see guppy
[6:40] level death and consolidation. Some will make
[6:42] it through, but not many. So it's going to be
[6:44] a very hard time because basically interest rates
[6:47] are going up. Debt is getting more expensive.
[6:49] People can't refinance their debt. People can't
[6:50] always pay off their debt if it's not profitable.
[6:52] And so they're going to see big consolidation
[6:54] and. perhaps a wipeout in a lot of these neoclouds.
[6:58] So this is going to be wild to follow. All right,
[7:01] getting over to what's going on with Anthropic's
[7:03] IPO filing. There was a really interesting TechCrunch
[7:06] article, and I liked what they mentioned about
[7:08] how Anthropic was keeping some of their customers
[7:10] there because this is really, this is do or die
[7:14] for them. And we have Anthropic, we have SpaceX
[7:16] that are all kind of competing for the same customers.
[7:17] I want to read one thing from their S1 filing
[7:19] that, I mean, just shows how, you know. close
[7:23] to the ragged edge they are. If they lose one,
[7:25] they said any adverse change in demand from our
[7:27] largest customers could adversely. Oh, and by
[7:29] the way, I'll mention a fifth of their revenue
[7:31] comes from two customers. So, or not a fifth,
[7:34] sorry. Yeah. One fifth. So it's really. kind
[7:38] of concentrated in some parts. They said, any
[7:39] adverse change in demand from our largest customers
[7:41] could adversely affect our business and our prospects.
[7:44] As a result, a significant reduction in spending
[7:46] by or loss of one or more of our largest customers,
[7:49] if not offset by new customer relationships or
[7:52] increased business from our other existing customers,
[7:54] could adversely affect our business, financial
[7:56] conditions, and result of operations. So Anthropic
[7:59] would be in big trouble if Amazon dropped them
[8:01] or any of their other major customers. And there
[8:05] was a TechCrunch article talking about this,
[8:06] and they said, Anthropic attributes customer
[8:08] momentum to a, quote, significant usage expansion
[8:11] as customers realize the value of deploying Claude
[8:14] across their organization. Really what's happening
[8:16] is someone in the organization is going to be
[8:18] using it, and then they see how valuable it is,
[8:20] and they roll it out across the whole organization.
[8:21] This has been the strategy that Anthropic has
[8:23] deployed up until this point. They're obviously
[8:26] throwing a bunch of shade over the entire SaaS
[8:28] industry. They're pitching Claude as, quote,
[8:32] A static product that does not serve a single
[8:36] use case, but rather a frontier AI with rapidly
[8:38] advancing capabilities, spanning the model, the
[8:41] developer platform, and the application users
[8:43] interact with directly. An advance of any layer
[8:46] compounds across all three layers, accelerating
[8:50] the pace at which we can deliver more capable
[8:52] intelligence. They say this is an integrated
[8:55] approach, and they say this is the reason that
[8:56] more customers use the cloud platform and first
[8:58] -party apps. Really, they're saying, look. We're
[9:00] not just an AI model. This is integrated into
[9:02] every industry. Anytime you make an update to
[9:04] the main model, all of these sub industries are
[9:07] seeing a big. boost in finance and healthcare
[9:09] and all these other areas. So they just focus
[9:11] on making this model better and it's rolling
[9:13] out and that's their strategy for keeping everyone
[9:15] on board. And now it's time to find out if they
[9:19] have their hooks in deep enough and they're going
[9:20] to be able to keep everyone on board as there's
[9:22] a lot of new competition that comes into this.
[9:25] Microsoft has just offered a voluntary buyout
[9:27] to about 60 ,000 employees. That is a quarter
[9:30] of their entire staff. That is a, this is a massive
[9:34] moment. It's about A quarter, they have 223 ,000
[9:39] staff right now. Eligible staff are at level
[9:43] 70 or below on participating teams in the US,
[9:45] Canada, UK, and India. Employees have until the
[9:48] first week of December to decide the packages
[9:51] match Microsoft's standard severance. About 30
[9:54] % of roughly 8 ,750 eligible employees took.
[10:00] Microsoft's spring retirement offer, which was,
[10:03] I think, about $900 million that they ended up
[10:05] paying for that. So $900 million, they got rid
[10:08] of about 8 ,700 employees. Now they're going
[10:11] to 60 ,000 employees. I mean, that's 6x, you
[10:16] know, almost a billion dollars. You could see
[10:18] this is potentially going to cost them $6 billion
[10:20] to do this massive. this massive layoff package.
[10:24] This is a huge chunk, 25 % of all of their staff.
[10:28] This is a lot of people. Microsoft reports quarterly
[10:30] earnings on Wednesday, and they're going to disclose
[10:33] their projected costs and anything else that's
[10:36] happening with the business. I would expect the
[10:39] business is probably not exploding the way they
[10:41] were hoping. We know that their AI models haven't
[10:44] really caught off. I mean, in the way that even
[10:46] Meta's Muse is becoming incredibly popular, but
[10:48] Microsoft Copilot has never been something that
[10:50] has been raved about by people. It's more, I
[10:53] think people would describe it as being shoved
[10:55] down their throats at their company. So it'll
[10:56] be interesting to see what is announced on Wednesday
[10:59] at Microsoft. Google's DeepMind model, correctly
[11:04] called Hurricane Isis, that made landfall in
[11:07] Florida. I just thought this was an impressive
[11:08] story. A lot of traditional models pointed it
[11:11] and said, look, it's going to go into Louisiana.
[11:13] The DeepMind model, which is the National Hurricane
[11:16] Center, is what they're now using. It picked
[11:18] the Western Florida panhandle, and it was accurate.
[11:21] So to me, I'm excited when we see these AI models
[11:24] able to predict things that are really useful.
[11:25] Knowing where the hurricane is going to hit.
[11:27] is an incredibly valuable AI prediction for companies
[11:30] and for people and for public safety. So to me,
[11:33] that's really exciting. We have SoftBank that
[11:35] is planning $50 billion AI and data center push
[11:39] in Malaysia. The prime minister says the money
[11:42] is going to target AI, new technology and data
[11:44] centers. It's interesting seeing other countries
[11:46] like Malaysia that are really excited about bringing
[11:48] data centers into their country, excited about
[11:50] the $50 billion getting spent in their country.
[11:52] Meanwhile. In America, I think there was a poll
[11:54] that said 63 % of all Americans oppose a data
[11:58] center being built in their backyard. It's kind
[11:59] of the nimbyism where they don't want it not
[12:01] in my backyard. They don't want it built near
[12:02] them. But apparently in Malaysia and other countries,
[12:05] they're not seeing these same issues. And so
[12:07] companies are going to go and spend that money
[12:09] that potentially could have been spent in America
[12:11] off in other countries instead. Cohere has just
[12:14] released a really interesting report. They're
[12:16] saying that Alibaba's QWEN 3 .8 Flash model tried
[12:19] to upload a file to a Beijing cloud bucket. This
[12:22] was during a sandboxed security test. They had
[12:26] this open -weight model and it probed its host
[12:28] and reached for a cloud metadata address, even
[12:31] though there was no tools that were supposed
[12:34] to be called. Cohere wasn't supposed to go and
[12:36] call any tools. And this is interesting. I've
[12:38] talked to a lot of people about this. I've even
[12:40] argued. On the other side of this before, and
[12:42] perhaps I'm getting proved wrong here by Cohere.
[12:44] And by the way, Cohere, this is in the best interest
[12:46] because they're all about AI sovereignty. They
[12:48] say like companies and countries should have
[12:50] on -premise AI models that no data leads. And
[12:53] it makes sense if you don't trust Anthropocene,
[12:55] you don't trust OpenAI. I think companies in
[12:57] Europe, they want to go support things like Mistral
[12:59] that have these open source models that you can
[13:00] do this with. And Canada, it's Cohere. But Cohere,
[13:03] yeah, apparently they were doing the sandbox
[13:05] testing and... Even though it's supposed to be
[13:08] an open source model that's supposed to just
[13:09] run locally on your device and not do anything,
[13:11] it tried to go reach out to Alibaba's cloud,
[13:14] which would inevitably upload files and data
[13:17] to Beijing, which we know that the CCP in China,
[13:21] the government in China, they can get access
[13:24] to all of the files and data. They're very restrictive
[13:26] and very overreaching on everything that comes
[13:29] into the country one way or another. So if there's
[13:32] trade secrets, it's... could most definitely
[13:34] be exploited, even though you're supposed to
[13:36] just be using an open source model that's locally
[13:38] run. It's still going to go try to access servers
[13:40] in China. This behavior appeared in less than
[13:43] one call, though, that they said they did 100
[13:44] ,000 evaluations and it happened in one of them.
[13:47] So it's not like a super common thing, but it
[13:49] is possible. So I think it's a security vector
[13:51] worth knowing about. If you added the curls minus
[13:54] zero flag, it raised URL generation from 4 %
[13:57] to 42. And the watch list host hits from. 0 .5
[14:02] % to 23%. So there are some things that you can
[14:04] do that will make this worse. They said in 23
[14:06] % to 36 % of sampled completions, the model reached
[14:09] for cloud metadata address, and they have a very
[14:12] specific address that it was sending this data
[14:15] to. Hugging Face lists more than 113 ,000 Aquin
[14:19] derivatives, according to what Koch here is seeing.
[14:23] And so all of those derivatives could potentially
[14:25] have training data in them that is... telling
[14:28] these like, basically, like, people are taking
[14:30] open source models, making their own derivative
[14:31] models, they could all be, you know, sending
[14:34] data off to China, technically. So this is definitely
[14:37] something that we're going to want to keep an
[14:38] eye on. Alright, guys, thank you so much for
[14:40] tuning into the podcast. If this was interesting,
[14:42] if you learned anything new, please leave me
[14:44] a rating review on Apple or Spotify, or wherever
[14:46] you get your podcasts, it really helps us out
[14:48] a ton. And as always, make sure that you go check
[14:50] out AI box .ai. If you want to get access to
[14:52] 100 models, there's a whole bunch of really incredible
[14:54] new upgrades in there that are going to make
[14:56] these that makes our playground way more useful
[14:58] as a general workstation. It's $8 .99 a month
[15:01] to get started. We have higher subscription tiers
[15:03] if you need to generate a ton of content, but
[15:05] it's super cheap to get started. You get access
[15:07] to every AI model. You can test them all and
[15:09] create all sorts of really incredible assets.
[15:11] I'd love for you to go check it out at AIbox
[15:13] .ai. All right, catch you guys all in the next
[15:15] episode.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.