300 Megawatts for Your AI Agents: Why Anthropic's Compute Deal Boosts Reliability for the Mittelstand

If you've run into an error message on Claude Code or the Anthropic API in recent weeks, you were probably not alone. Between late May and mid-June 2026, reports of outages piled up - according to TechTimes, by June 16 it was already the tenth noticeable disruption within twelve days, with HTTP 500 errors and the dreaded HTTP 529 messages, where the system is fundamentally working but simply has no capacity left for new requests. For a company that relies on AI agents for its accounting, its reporting, or its customer service, that's not an abstract problem, it's a tangible operational risk: an AI agent that stalls right in the middle of month-end invoicing costs time, nerves, and, in doubt, trust in the entire automation effort.

It's worth taking a second look, in exactly this context, at a story that made the rounds again in early July 2026: Anthropic has secured the entire compute capacity of the SpaceX/xAI Colossus 1 data center. At first glance the news looked brand new - in fact, the deal had already been announced on May 6, 2026, and the coverage from July 7 simply revisited an announcement that was already roughly two months old. What makes this timing so interesting: on that same day, Anthropic rolled out noticeably improvements for users. That's no coincidence - it's evidence that more compute capacity translates directly into more usable capacity. What this concretely means, why the preceding outages weren't a one-off, and why the topic is especially relevant for the Mittelstand, we break down in this article along the verifiable facts - including the points that are still open.

The deal at a glance: 300 megawatts, 220,000 GPUs, one data center in Memphis

Colossus 1 sits in Memphis, Tennessee, in the Boxtown neighborhood - in a former Electrolux factory that xAI and SpaceX converted into one of the largest AI data centers in the world. With the deal from May 6, 2026, Anthropic secured more than 300 megawatts of power and over 220,000 Nvidia GPUs - a mix of H100, H200, and the newer GB200 accelerators. This isn't a share of the facility, but, according to consistent reporting, the entire available capacity of Colossus 1.

To put the scale in context: 300 megawatts roughly corresponds to the constant power demand of a mid-sized city. And the GPU mix is deliberately staggered - older H100 chips for stable, well-plannable inference loads, and the newer H200 and GB200 accelerators for especially compute-intensive requests and future model training. For you as a user, this essentially means: it isn't about a single hardware generation, but a broad, multi-year capacity reserve that can cover different load profiles - from simple chat to compute-intensive agent workflows.

On price: according to documents related to SpaceX's IPO, Anthropic reportedly pays around $1.25 billion per month, under a contract running through May 2029. For xAI and SpaceX, the deal could mean more than $40 billion in revenue over its full term. These figures don't come from an official Anthropic confirmation, but from reports on SpaceX's IPO filings - we're deliberately flagging them as "reportedly," not as officially confirmed fact.

It's also notable who becomes a business partner here: xAI, Elon Musk's company, is renting compute capacity to Anthropic - a direct competitor in the AI model market. For the reliability of your AI agents, this arrangement is ultimately secondary, but it shows just how far the competition for compute power now extends beyond classic competitive boundaries: whoever has spare capacity today rents it out - if need be, even to their own competitor, because industry-wide demand for compute exceeds supply.

Why compute capacity matters to you as a Mittelstand business owner

You might think gigawatt figures and data centers in Tennessee are far removed from a Mittelstand company's day-to-day business. The opposite is true. Every request your AI agent sends to a language model - whether it's checking an invoice, drafting an email, or analyzing a dataset - ultimately needs physical compute power in a data center. When that capacity gets tight, you feel it directly: as a delay, as an error message, or as a stricter usage limit in your contract. Unlike classic software licensing, where "more users" usually just means "more servers," scaling large language models is technically more demanding and capital-intensive - a single modern AI data center costs billions and takes months to years to build.

That became noticeable in early summer 2026. In addition to the outages already mentioned in June, there was a broader outage as early as June 2, 2026. Anthropic CEO Dario Amodei has publicly confirmed that compute capacity is currently the central bottleneck for Claude's model development and availability - not a lack of ideas or algorithms, but simply too few available chips for strongly growing demand. That's a remarkably candid statement from the CEO of a company that ranks among the leading AI providers: demand for Claude is growing faster than new data centers can be built.

Compute capacity is currently the limiting factor - not only for how capable new models become, but also for how reliably existing models are available day to day. Anthropic CEO Dario Amodei has publicly framed it this way, and the Colossus 1 deal is the direct corporate response to it.

For you as a decision-maker, this means: the stability of the AI agents you rely on doesn't depend solely on software quality, but decisively on how well the provider succeeds in providing sufficient physical compute power. That's exactly where the Colossus 1 deal comes in - and exactly why it's worth looking not just at model quality, but at the infrastructure strategy behind it, when choosing AI partners.

The proof: what changed immediately on May 6, 2026

What's really remarkable about this deal isn't just its size, but the timing tied to concrete improvements for users. On the very day the Colossus 1 deal became public, Anthropic rolled out several noticeable adjustments:

• The five-hour usage limits for Claude Code were doubled.
• Peak-load throttling was lifted.
• Opus API limits were significantly raised - according to reporting from 9to5Google, by as much as plus 1,500 percent on Tier-1 input token limits per minute and plus 900 percent on output token limits per minute.

That's an unusually concrete, verifiable link: more secured compute capacity, translated into higher usable quotas on the very same day. For companies that use Claude Code or the Anthropic API at scale - for example for automated evaluations, document processing, or AI-assisted software development - this isn't a theoretical promise, it's an already-realized, measurable improvement in usable capacity. Anyone who previously hit a usage limit multiple times a day could notice this difference directly in their own working environment.

The increase in Tier-1 limits in particular is practically relevant: smaller and medium-sized companies still early in their API usage often start on lower pricing tiers. When input and output token limits per minute rise this sharply precisely at that level, it's not just large enterprise customers who benefit, but also growing Mittelstand businesses that are just now ramping up their AI usage.

Important for context: these effects have already been visible since May and June 2026. According to industry reports, new infrastructure investments typically take 12 to 24 months to fully translate into real, available capacity. That means the positive effects of the Colossus 1 deal on stability and availability are likely to keep intensifying over the rest of 2026 and into 2027 - not as a one-time jump, but as a gradual improvement that keeps building month by month.

More than one deal: Anthropic's comprehensive infrastructure strategy

Colossus 1 is an important, but by no means the only, building block in Anthropic's capacity strategy. In parallel, Anthropic has struck other major agreements:

• An expanded partnership with Google and Broadcom for around 3.5 gigawatts of TPU capacity, expected to be available from 2027.
• An agreement with Amazon for up to 5 gigawatts of Trainium capacity, tied to an investment volume of more than $100 billion over ten years.

In total, Anthropic now has around 10 gigawatts of contractually secured hyperscaler capacity with Google and Amazon - on top of the GPUs from the Colossus 1 deal. For you as a customer, this point is strategically important: the reliability of your AI agents doesn't hinge on a single data center or a single partner. A diversified portfolio of multiple infrastructure providers significantly reduces the risk of a single point of failure - if one provider or one region temporarily goes down, there's structurally more fallback capacity than there was a year ago. It's comparable to a company that no longer draws its power from a single power plant, but from several independent sources.

In addition, Anthropic has also signaled interest in a partnership with SpaceX for several gigawatts of orbital, i.e. space-based, AI compute capacity. That's expressly not yet a signed contract, but a statement of intent - we mention it here as an outlook on the long-term direction of thinking, not as a decided measure. It does show, however, that Anthropic treats compute capacity as a core strategic topic that goes well beyond the single Colossus 1 deal.

This investment offensive can also be assessed economically: according to industry reports, Anthropic is on track for annual revenue of around $47 billion and is aiming for profitability from 2029 - a year earlier than its main competitor, OpenAI. This economic development explains why such large infrastructure investments make sense for Anthropic in the first place, and underscores that this is a long-term strategy rather than a short-term reaction to individual outages. A company running on a tight budget in the short term doesn't simultaneously invest in Colossus 1, in Google/Broadcom TPUs, and in Amazon Trainium capacity.

What this means for your AI agents day to day

Let's translate the numbers into your everyday operations. Many Mittelstand companies deploy AI agents precisely where recurring load spikes happen: overnight batch processing of accounting data, month-end closes, quarterly reporting cycles, or billing runs. Exactly these scenarios - many requests in a short time, often outside regular business hours - were the ones most affected by the capacity bottlenecks and the HTTP 529 errors mentioned earlier in recent months. Picture an agent that checks and categorizes hundreds of receipts every overnight run: if the connection drops right at that point, nobody shows up the next morning to finished results, but to a half-processed queue.

The measures rolled out on May 6, 2026 address exactly these situations directly: doubled usage limits on Claude Code, lifted peak-load throttling, and significantly higher API rate limits. So if your company, for example, uses an AI agent for automated invoice checking, for preparing month-end closes, or for analyzing large volumes of data, you benefit from precisely the improvements directly tied to the Colossus 1 deal. Use cases like automated recruiting, content production, or research agents that make many requests in a short time also run noticeably more reliably as a result.

A realistic assessment matters here too: it would be an exaggeration to claim Anthropic was previously "constantly down." It's more accurate to say that certain high-load phases saw an elevated error rate - a pattern many companies recognize from their own fluctuating IT loads, such as Black Friday weekends in e-commerce or year-end closing in accounting. The difference is that Anthropic is responding with one of the largest single investments in compute capacity any AI provider has made public to date, complemented by a broadly diversified portfolio of further capacity sources. That's a different pattern than mere crisis management - it's forward-looking capacity planning.

Realistic expectations: what's still open

For all the positive framing, an honest assessment also has to note what the Colossus 1 deal does not guarantee. There's currently no verified statement from Anthropic that API prices will fall, or even just stay stable, because of the added capacity. Falling or stable prices are a plausible medium-term expectation, but not yet a confirmed fact - and shouldn't be communicated as one. Anyone investing in AI agents today should still base their calculations on current terms and treat possible price cuts as a positive bonus effect, not as a fixed planning assumption.

Likewise, Colossus 1's current utilization can't be verified in real time as of today; the solid facts date from May and June 2026. And the widely cited "tenth disruption in twelve days" is based on a single industry report, not on an official Anthropic status page - we're deliberately flagging it as a report, not as conclusively confirmed fact. The intent signaled by SpaceX/xAI regarding orbital compute capacity is also, as mentioned, not yet a signed agreement and shouldn't factor into concrete capacity planning accordingly.

What can be said with high confidence: the direction is right. More secured compute capacity, immediately followed by higher usable quotas, embedded in a broadly diversified infrastructure strategy spanning multiple providers - that's a solid foundation for more reliability through the rest of 2026 and beyond. For companies still hesitant to deploy AI agents in business-critical processes, this removes a central counterargument: the fear of unpredictable outages at the wrong moment.

In the end, what matters to you as a Mittelstand business is less the exact gigawatt figure of a data center in Memphis, and more the question of whether your AI agents run on infrastructure that's visibly investing in the right direction. Based on the facts at hand, that question can be answered with a yes - with the caveats noted around pricing and exact capacity levels. Whoever invests now in well-planned, robustly connected AI agents benefits both from the improvements already realized and from the further effects likely to unfold over the coming months and years.