Meta to start manufacturing ‘Iris’ AI chips in September, targeting 14GW in 2027
An internal memo reviewed by Reuters says Meta will put its in-house data center chip into production as it races to lower AI costs — alongside a newly priced coding model for developers.
One-Line Summary
Big Tech moves to cut AI costs and speed adoption — Meta starts in-house chips and undercuts coding-model prices while OpenAI buys services talent to embed with enterprises.
Big Tech
Meta to start manufacturing ‘Iris’ AI chips in September, targeting 14GW in 2027
Meta, which operates Facebook and Instagram, plans to put an in-house data center AI chip into production starting in September as part of a four-generation Meta Training and Inference Accelerators (MTIA) program; the internal memo reviewed by Reuters says the effort aims to boost overall computing capacity to 14 gigawatts in 2027. The chip is code-named “Iris.” 1
Early testing lasted six weeks and found no major issues, according to the memo. Meta tailored the silicon for its own needs and is working with Broadcom on design and TSMC to manufacture it — a path Reuters says is likely to lower computing costs and reduce dependence on Nvidia and AMD. 1
The chip is meant to augment, not replace, Meta’s large fleets of GPUs, and the company has lined up deals for memory, flash storage, and fiber optics to support the rollout. Reuters notes the completion of bug testing and the production timing had not been previously reported — a sign of renewed momentum after earlier in-house chip efforts struggled. 1
SpaceXAI launches Grok 4.5 with lower token prices
SpaceXAI released Grok 4.5 as a general-purpose “workhorse” model for coding, clerical work, research, and writing; Elon Musk described it as an “Opus-class model” and said internal assessments find it roughly comparable to Opus 4.7 but faster. 2
TechCrunch reports Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, contrasting it with Anthropic’s Opus 4.7 at $5/$25 and OpenAI’s tiers ranging from $1/$6 to $5/$30 depending on the version. If real-world performance matches the claims, the pricing could pressure costs for teams doing high-volume inference. 2
Industry & Biz
OpenAI’s deployment arm to acquire Northslope
OpenAI’s Deployment Company — launched in May to help enterprises implement AI in core operations — agrees to acquire Northslope, an applied AI firm, marking its second enterprise-focused acquisition after Tomoro. Axios reports the unit, majority-owned and controlled by OpenAI, began with $4 billion to fund acquisitions; terms are undisclosed and the deal is subject to customary regulatory approvals. 3
Axios says the acquisition expands a bench of hundreds of “forward deployed engineers” who embed with customers to build systems — signaling that labs are competing on implementation know-how as much as on model performance. Anthropic is also building an AI services company aimed at mid-sized businesses. 3
New Tools
Meta releases Muse Spark 1.1 for coding agents, sets $1.25/$4.25 pricing
Muse Spark 1.1 is a multimodal AI model designed for agentic coding and complex workflows; Meta is opening access via a new Meta Model API in public preview for U.S. developers and making it available in “Thinking” mode in the Meta AI app and on the Meta AI website, with $20 in free credits per new API account. 4
CNBC reports the model’s pricing at $1.25 per million input tokens and $4.25 per million output tokens, which Meta characterizes as “very aggressive and attractive” compared with similar offerings. The company is positioning Spark 1.1 as its strongest model yet for agentic and coding work. 5
CNET highlights added capabilities like orchestrating multi-agent systems, improved computer-use workflows, and actively managing a 1‑million‑token context window to handle longer tasks and code migrations. Together, the launch and pricing suggest Meta aims to compete on both capability and total cost of ownership. 6
Google Photos adds AI ‘Video Remix’ for fast video edits
Google Photos is rolling out “Video Remix,” an AI feature powered by Gemini Omni that can transform videos in seconds — including cinematic relighting, background swaps, and artistic styles — accessible from the Create tab. 7
The feature starts rolling out to eligible Google AI Plus, Pro, and Ultra subscribers across specified countries (including South Korea), reflecting Google’s push to embed generative video editing into everyday consumer tools. 7
Community Pulse
Hacker News (293↑) — Mixed reaction: some want openness (even datasets), others scrutinize Muse Spark’s pricing versus rivals. 8
"the way he could really be the spoiler king is to release an their training dataset to open source… doubt he’d go that far." — Hacker News 8
"The cached input pricing is a good ratio. Compare with Grok 4.5 which came out at $2/$6 but then quietly charges $0.50 per 1M cached input tokens. That's as high as Opus 4.8!" — Hacker News 8
What This Means for You
Infrastructure costs are becoming a competitive lever. Meta’s in-house “Iris” chip program signals more control over cost and supply — pressure that could eventually show up as lower per-token prices or faster features in the apps and services you use. For tech buyers, it’s a reminder to revisit pricing with vendors as chip supply and efficiency improve. 1
On tools, Muse Spark 1.1’s public preview and pricing ($1.25 input / $4.25 output per 1M tokens, plus $20 credits) give teams a concrete new option for code automation, migrations, and agentic workflows. If you’re experimenting with AI coding, run the same task across your current model and Spark 1.1 to compare cost, speed, and success rate. 5
Competing price points like Grok 4.5’s $2/$6 underscore that token economics now determine whether an AI use case scales. Before a broader rollout, estimate monthly token spend for your top 2–3 tasks to avoid surprise bills and to pick a model tier aligned with your workload. 2
Enterprise adoption is shifting from “model choice” to “who implements with you.” OpenAI’s Northslope acquisition adds embedded engineers who can build systems alongside your teams — helpful if internal capacity is thin but you need production-grade deployments with security and governance. Consider scoping an info session with your vendor’s services arm. 3
Action Items
- Try Muse Spark 1.1 in Thinking mode: Open the Meta AI app or meta.ai and run a small bug-fix or refactor task to gauge quality and latency.
- Apply for the Meta Model API preview: If you’re in the U.S., sign up and use the $20 in credits to measure token costs on one real workflow.
- Recalculate your AI coding budget: Compare Muse Spark 1.1’s $1.25/$4.25 with Grok 4.5’s $2/$6 for one weekly task to see potential savings.
- Request info on OpenAI deployment services: Ask your OpenAI account contact how the Deployment Company/Northslope team works with customers and what a security/legal review would require.
- Prototype a marketing clip with Video Remix: If you’re a Google AI Plus/Pro/Ultra subscriber in an eligible country, use Google Photos’ Video Remix to relight and style a 15–30s clip for your next post.
Comments (0)