Sources: Microsoft is in talks to invest $10B in OpenAI at a $29B valuation, taking 75% of OpenAI's profits until the investment is recouped and then owning 49%

Microsoft has been in talks to invest $10 billion into the owner of ChatGPT, the wildly popular app that has thrilled casual users …

Semafor 2023-01-10

Context & Ripple Effects

The proposed terms would deepen a relationship already built through Microsoft’s 2019 and 2021 investments; the reported deal was subsequently described as a $10B investment spread over years. Later coverage shows that the initial profit-and-equity framework did not settle the partnership permanently: the companies returned to negotiations over Microsoft’s equity as OpenAI pursued a for-profit structure.

The importance is not just the size of the financing but its economics. Microsoft would receive both a near-term claim on profits and a large eventual ownership position, tying OpenAI’s funding model closely to one strategic backer as its valuation and capital needs grew.

First-order effects

OpenAI would gain access to a reported $10B commitment, while Microsoft would secure 75% of profits until repayment and a reported 49% ownership stake afterward.
The arrangement gives Microsoft an immediate economic claim on OpenAI’s commercial upside rather than a conventional minority-investment exposure.

Second-order effects

OpenAI’s subsequent outside fundraising and corporate-structure decisions would have to accommodate Microsoft’s profit share and prospective stake, a tension reflected in later fundraising talks at a $100B-plus valuation.
Microsoft’s financial exposure gives it a stronger incentive to connect OpenAI’s growth to its own platform and infrastructure business; the later Azure purchasing commitment illustrates how capital and service demand became linked.

Third-order effects

If frontier labs continue to fund compute-intensive development through strategic investors, ownership, revenue rights, and infrastructure purchasing will increasingly be negotiated as one package rather than as separate transactions.
The pattern concentrates bargaining power among the companies able to provide both large-scale capital and underlying AI infrastructure, while forcing labs to repeatedly renegotiate governance as their commercial value rises.

The trend: Frontier AI funding is evolving into infrastructure-linked strategic finance, where investment terms shape both ownership and the buyer-supplier relationship.

Discussion

@jamespmcleod James McLeod on x
OpenAI has gotten a LOT of PR mileage over the years about its roots as a nonprofit and the idea that its profits are “capped” and its goals are fundamentally altruistic. Anyway, watching it fall into the loving arms of Microsoft fills me with a lot of cynical feelings. https://t…
@marypcbuk Mary Branscombe on x
Open AI has more reputation than profit and all that Azure compute is expensive... https://twitter.com/...
@bdsams Brad Sams on x
Coupled with the Activision acquisition, Microsoft is looking to spend $80b in 2023 on investments for its future. https://twitter.com/...
@lizrhoffman Liz Hoffman on x
This deal is an odd one. Microsoft gets 75% of OpenAI's profits until it gets its $10bn back (unclear whether cloud-computing fees count toward that but would certainly speed things up.) Then it flips to reflect the fact that MSFT would own half this thing https://www.semafor.com…
@carnage4life Dare Obasanjo on x
Satya Nadella cements his entry in the CEO hall of fame by spotting OpenAI early and locking down access to its fruit for use in Office and Bing. FTC is worried about Activision which is gaming's when Microsoft is already locking down the future of AI. https://www.semafor.com/...…
@silviakillings Silvia Killingsworth on x
Clippy's revenge https://www.bloomberg.com/...
@reedalbergotti Reed Albergotti on x
Scoop with @lizrhoffman: With a $10 billion investment, Microsoft could soon double down ( tenfold down?) on OpenAI, maker of ChatGPT and other mind blowing AI tools. Microsoft would also get 75% of all OpenAI profits for a time. https://www.semafor.com/...
@davemcclure Dave McClure on x
um... wow. 😳😱💥 https://twitter.com/...
@bentossell Ben Tossell on x
The real AI winner: Microsoft - $1bn in OpenAI (2019) - Owns GitHub which released Copilot - Dalle powers Bing Image Creator - Valle which produces an AI voice from 3s of input - GPT in Word, PowerPoint, Outlook etc - GPT-powered Bing - $10bn, 49% stake in OpenAI + 75% profit
@lizrhoffman Liz Hoffman on x
We asked ChatGPT to write a version of our story from the viewpoint of a Wall Street analyst and, well, it is superb. https://www.semafor.com/... https://twitter.com/...
@nonmayorpete Pete on x
The Microsoft <> OpenAI news is much bigger than we first thought: - $10 billion investment in the works, with Microsoft getting 75% of revenue until it recoups all $10b - GPT in Office, not just Bing
@hammer_mt Mike Taylor on x
How much did Microsoft already own then? Must be ~20% if they will eventually own 49% and they're buying 1/3rd today? https://twitter.com/...
@lizrhoffman Liz Hoffman on x
Submit your best Clippy jokes https://twitter.com/...
@ericabrescia Erica Brescia on x
I don't have any insider info, but would put money on this being true. Another brilliant move by Microsoft if they pull it off. https://www.semafor.com/...
@omerperchik @omerperchik on x
When Microsoft is playing Mr. Wonderful @kevinolearytv https://twitter.com/...
@levie Aaron Levie on x
If real, this deal-making by Microsoft is pretty incredible to watch. https://www.semafor.com/...
@seyitaylor @seyitaylor on x
Interesting that MSFT is front-loading the profit share. Wonder if we see more of these with SV startups. https://twitter.com/...
@dsmart Derek Smart on x
Well, they did pay Notch $2.5B for Minecraft, so this seems legit. 😂😂😂 https://www.semafor.com/...
@tomwarren Tom Warren on x
Microsoft rumored to invest $10 billion in OpenAI. The investment in the ChatGPT creator would see Microsoft net 75% of profits and a 49% stake in OpenAI https://www.theverge.com/... https://twitter.com/...
@nonmayorpete Pete on x
Where the $10b investment sits relative to Microsoft's largest *acquisitions*: 1. Activision Blizzard ($68.7b, pending) 2. LinkedIn ($26.2b) 3. Nuance ($19.7b) 4. OpenAI $10b investment 5. Skype ($8.5b) 6. ZeniMax Media ($8.1b) 7. GitHub ($7.5b)
@mikeisaac Rat King on x
Microsoft really, really wants in on Open AI and Chat GPT https://www.semafor.com/...
@simon_ree Simon Ree on x
The rumours of the imminent death of $GOOGLE search are only going to accelerate Microsoft Considers $10 Billion Investment in ChatGPT Creator https://www.bloomberg.com/...
@borrowed_ideas @borrowed_ideas on x
Google paid ~$650 Mn to acquire DeepMind in 2014. OpenAI is currently “valued” $29 Bn. OpenAI's revenue is “tens of millions”. DeepMind's 2020 revenue was $1.13 Bn and was profitable. (although Google is probably the main customer). What would DeepMind be “valued” today?
@tedgioia Ted Gioia on x
Microsoft is making a big bet that it can overtake Google with AI-driven search—this is rapidly turning into the most high stakes battle in tech. https://twitter.com/...
@martinsfp @martinsfp on x
Microsoft could invest as much as $10 billion in ChatGPT and DALL-E maker OpenAI, in a multi-year deal. It was reported over the weekend that Microsoft is looking at putting GPT into apps like Word and Outlook. https://www.bloomberg.com/...
@trungtphan Trung Phan on x
Microsoft looking at a $10B investment into OpenAI: • At a $29B valuation • MSFT gets 75% of OpenAI profits until it recoups $10B • After recouping $10B, the ownership structure becomes MSFT (49%), other investors (49%) and OpenAI non-profit parent (2%) https://twitter.com/...
@evanwolf@mastodon.social Phil Wolff on mastodon
#Microsoft's VALL-E seems to be several orders of magnitude faster at getting text-to-speech to sound like someone's voice (three seconds). There are sentimental uses, like making a voice of someone you mourn. …
@thecarteldel @thecarteldel on x
Microsoft have announced their AI “VALL-E” Using a 3-second sample of human speech, it can generate super-high-quality text-to-text speech from the same voice. Even emotional range and acoustic environment of the sample data can be reproduced. Here are some examples. https://twit…
@drjimfan @drjimfan on x
DALL-E generates pixels from text. Now meet its cousin, VALL-E, that generates audio from text @MSFTResearch! VALL-E's resemblance to DALL-E v1 and Parti @GoogleAI is striking. Image and audio are both continuous signals, but they can be quantized into discrete tokens. 1/🧵 https:…
@sytelus Shital Shah on x
Just about to wrap up my day and saw VALL-E! Wow!! This model takes 3 seconds of speech sample for a person and can synthesize text-to-speech in same voice with unbelievable fidelity. It can maintain even emotion and acoustic environment in the sample. https://valle-demo.github.i…
@glenngabe Glenn Gabe on x
What could go wrong? :) BTW, Adobe created this years ago, but never released it... -> Microsoft unveils VALL-E, a text-to-speech AI model trained on 60K hours of English speech that can simulate a person's voice from 3 seconds of sample audio https://arstechnica.com/... https://…
@austin_walker Austin Walker on x
This comes days after a big AI audiobook announcement from Apple, part of a trend (which Google is also part of) that voice actors immediately noted was a risk to both their own stability and to the field in whole, as audiobooks serve as a good place for junior talent to break in…
@shobz Shoaib Taimur on x
“Perhaps owing to VALL-E's ability to potentially fuel mischief and deception, Microsoft has not provided VALL-E code for others to experiment with, so we could not test VALL-E's capabilities.” https://arstechnica.com/...
@wbm312 Whitney Merrill on x
“my voice is my passport” https://twitter.com/...
@ejwalters Ed Walters on x
This is the (quietest) third leg of the generative AI stool. GPT-3 for text, DALL-E/Stable Diffusion for images, and VALL-E for mimicking voice. Any words in the voice of any person. Deep fakes are about to make people question news, facts, and evidence. https://twitter.com/...
@counternotions Kontra on x
Text-to-speech? The other shoe drops: https://twitter.com/...
@amogh42 Amogh Vaishampayan on x
It can maintain the acoustic environment of the speaker pretty well. I expect this to be very be useful in movies and shows where someone is speaking over the radio or a phone. https://twitter.com/...
@emmalbriant @emmalbriant on x
Concerning new advances in deception here... https://twitter.com/...
@drjimfan @drjimfan on x
OpenAI Whisper (speech2text) is trained on 680K hours of audio. VALL-E is only trained on 60K hours. This means the potential of text2speech models is way under-tapped. OpenAI can just scale up VALL-E 10x on their private training set and nail it. It's such a low-hanging fruit. h…
@amogh42 Amogh Vaishampayan on x
VALL-E also synthesizes voice variations for the same input text. https://twitter.com/...
@poppy_northcutt Poppy Northcutt on x
I hope our legislators are paying attention to these developments cause definitely need legislation to prevent the abuse that is sure to come of this. @VoteAnnJohnson @RepFletcher @whitmire_john https://arstechnica.com/...
@suhail @suhail on x
An AI breakthrough every single week. TikTok for singing/rap incoming... https://mpost.io/...
@gaurabc Gaurab Chakrabarti on x
Absolutely terrifying. How will we know if anything is real in the digital realm? https://twitter.com/...
@chrismessina @chrismessina on x
Now anyone can have James Earl Jones (or anyone else for that matter!) voice their radio ads. Also relevant given @AppleBooks's just-announced Digital Narration audiobooks. https://www.producthunt.com/ ... #VoiceComputing #SyntheticMedia https://twitter.com/...
@drjimfan @drjimfan on x
Here's the recipe to make Siri/Alexa 10x better: 1. Whisper to convert speech to text. Best open-source speech model out there. 2. ChatGPT to generate smart home API calls and/or text response. 3. VALL-E to synthesize speech. It can mimic anyone's voice sample! Quick figure 1/3 h…
@drjimfan @drjimfan on x
Let's face it, Siri sucks. For anyone out there who wants to build Siri v2.0, you have my blessings 🙌! I look forward to it in 2023. Whisper: https://openai.com/... VALL-E: https://arxiv.org/.... Still waiting for Microsoft to release the models. See my deep dive thread: https://…
@altryne @altryne on x
VALL-E is quite quite impressive! I wonder if it can be trained on other languages, but getting a very short (3 seconds) or speech and making it into a synthesized voice sounds amazing! https://twitter.com/...
@ralphbrooks Ralph Brooks on x
@sytelus The ability to preserve emotion is definitely the cool part. Makes me wonder to what extent it will reduce the time needed to turn books into professionally recorded audiobooks. https://twitter.com/...
@drjimfan @drjimfan on x
Note that VALL-E is similar to DALL-E v1, NOT v2. DALL-E v2, Stable Diffusion, and Imagen rely on diffusion to regress continuous signals, instead of predicting discrete code sequences. 7/
@thecarteldel @thecarteldel on x
At the moment VALL-E can only read, not necessarily PERFORM with the emotional, tonal and pacing range of a voice actor. However, much of the audiobook industry relies on a lot of junior voice actor talent that will undoubtedly feel the brunt of this first.
@daftlimmy @daftlimmy on x
VALL-E https://twitter.com/... https://twitter.com/...
@shobz Shoaib Taimur on x
This is scary as it can be used by bad actors to sow the seeds of chaos. https://twitter.com/...
@schachin Kristine S on x
.@Microsoft oh yes let's please make something that people can use to imitate people's voices as though disinfo isn't bad enough. Just cuz you can doesn't mean you should and you're starting to get very scary esp what you're releasing to the public. https://arstechnica.com/...
@drjimfan @drjimfan on x
Here's the recipe to make Siri/Alexa 10x better: 1. Whisper to convert speech to text. Best open-source speech model out there. 2. ChatGPT to generate smart home API calls and/or text response. 3. VALL-E to synthesize speech. It can mimic anyone's voice sample! Quick figure 1/3 h…
@steventey Steven Tey on x
Surprised there isn't more chatter around VALL-E This new model by @Microsoft can generate speech in any voice after only hearing a 3s sample of that voice 🤯 Demo → https://valle-demo.github.io/ https://twitter.com/...

Chronicles