Parameter Update: 2026-21

Still no new OpenAI model in sight, but at least WWDC is happening this week. Last week was surprisingly quiet all things considered, though some of the finance stuff feels like playing with fire?

Microsoft Build

If you've been following Microsoft closely since their public semi-breakup with OpenAI, you know they're now set on becoming a full model company themselves. Build was a rather big update on these plans. In total, Microsoft announced five new models worth paying attention to:

• MAI-Thinking-1: sparse MoE with 35B active at roughly 1T total parameters. 256k context window, but otherwise tuned like a proper current-gen model. Microsoft claims users preferred it to Sonnet 4.6 in side-by-side evaluation, but benchmarks are a slightly more mixed bag.
• MAI-Code-1-Flash: Much smaller model at 5B active, tuned for coding. In their blog post they compare it to Haiku 4.5, so definitely not a big model contender.
• MAI-Image-2.5: Taking #4 in the image generation arena, this one actually beats Nano Banana Pro (but, in the arena, so does Nano Banana 2, so not sure how much that matters). Still very impressive!
• MAI-Transcribe-1.5: Apparently "the best transcription model in the world", but I actually don't know enough about that space to make that call.
• MAI-Voice-2: Very realistic TTS model. Again, not too familiar with the space, but the demos definitely sound great.

Most of these aren't state of the art in any way that matter, but they are our best indication yet that Microsoft is serious about their model ambitions (even if they still feel mostly like a hedge against OpenAI). It's also, frankly, incredible that Suleyman has been able to build competitive models in all dimensions in the timeframe they've had. Not sure if Microsoft is willing to commit the ressources required to push the frontier, but it certainly feels like they at least have the option now. In their announcement blog, they are very explicitly framing these as the latest outputs of the "hill-climbing machine" that is their training setup, so let's see them continue to climb?

Finance Stuff

SpaceX IPO Update

The SpaceX IPO circus continues to be one of the more interesting finance stories in tech right now. Shares are set to go live at the end of the week (June 12th), so we're in full launch mode now.

As for the IPO itself: SpaceX is targeting a June 11 share sale / pricing and June 12 Nasdaq trading debut, under ticker SPCX, with a reported target raise of roughly $75B at around a $1.75T valuation. The prospectus is now public, and hilariously, starts of with roughly 15 pages of pictures of rockets and satellites.

In my timeline, people are still convinced this IPO is mainly designed to provide exit liquidity for existing SpaceX shareholders, while on the other side of the table, people are increasingly pitching the company as an AI infrastructure play. With this week's multi-year cloud services deal with Google where Google pays $920M per month from October 2026 through June 2029 for access to roughly 110,000 Nvidia GPUs plus related compute, and the already announced Anthropic deal (together totallying over $70B in aggregate if they run to term), that pitch is not completely off-base?

Either way, we'll know more by the end of the week, so stay tuned, I guess.

Google Equity raise

The other giant finance story this week: Alphabet is raising equity.

Technically, Alphabet first announced an $80B equity capital raise to expand AI infrastructure and compute. The structure was: $30B in underwritten public offerings, $40B through an at-the-market program expected to begin in Q3, and a $10B private placement to Berkshire Hathaway. The stated use of proceeds is mostly AI infrastructure / compute.

The “why” is pretty straightforward: AI capex has gotten insane. Google says demand for its AI solutions is exceeding available supply; it guided 2026 capex to $180–190B and said 2027 capex would increase significantly again. It also disclosed that it had generated $174B in operating cash flow over the prior twelve months, while raising more than $85B of debt over the past year, taking total debt above $100B.

I don’t think the interesting question is “can Google afford this?” They probably can. The interesting observation is that we are watching the largest tech companies become capital-intensive infrastructure companies in real time - and Google decided to do so on external funds. The old software-company fantasy was infinite margins, tiny distribution costs, and buybacks forever. The new frontier AI fantasy is: build another hundred billion dollars of datacenters, raise equity if needed, and hope the models are good enough that this doesn't kill you before we get to superintelligence. Let's see how that plays out.

Anthropic IPO filing

IPO season is officially in full swing. While prediction markets were expecting OpenAI to go first, this week, Anthropic announced they confidentially submitted its draft S-1 filing to the SEC this week. The company did not disclose the number of shares, pricing, valuation, or exact timing, but the filing gives it the option to go public after SEC review, depending on market conditions.

This comes right after Anthropic’s absurd $65B Series H at a $965B valuation, and, fingers crossed, might give us a great glimpse at frontier lab financials. At least people on Twitter are being totally normal about it.

New image generation models

Besides Microsoft's new image model, there were two more new launches worth talking about:

Ideogram 4

Ideogram is still around, and this week dropped their first open weights model. At #9 in the arena it comes in just below Nano Banana Pro, but again, take these results with a grain of salt. The fun technical detail is that the text encoder is Qwen3-VL-8B-Instruct in text-only mode, and the model is trained around structured JSON captions with per-element styling, bounding boxes, and color palettes. The nf4 checkpoint fits on a single 24GB GPU, which makes this probably the best self-hostable image model out there right now.

Reve 2.0

The potentially more significant image model this week comes from Reve. Their core bet is that image generation needs an intermediate representation: a structured “layout” where every element has a location, size, local description, and optional references / colors. In their framing, layout is to images what HTML/SVG is to webpages. The model can generate from layouts, users can edit layouts directly, and agents can reason over them. If nothing else, that approach resulted in #2 in the image gen arena, one of the more visually appealing blog posts I've read in months, and one of the sickest looking image-gen animations in recent memory (plus extremely good steerability).

Reve 2.0 is "the best image generation model made by a sub-$1T company, trained on 10x fewer GPUs", which is a very funny benchmark category but also probably directionally useful - not everyone has access to OpenAI-level compute.

Instagram Support Agent

Finally, this weeks security nightmare: Meta gave an AI support agent the ability to modify Instagram account recovery details, and attackers used that to hijack accounts by convincing the bot to attach a new email address to target accounts. Once the attacker-controlled email was linked, they could trigger the normal password reset flow. Affected accounts included high-profile handles, including the Obama-era White House Instagram account, Sephora, and others.

If your agent can change the recovery email on an account, then your agent is not a chatbot. It is an identity and access management system with a language model stapled to the front. That should probably come with slightly more friction than “lmao sure”.