Confidential offering, West Midtown Atlanta
1000 Hemphill is an operating private data center with its own carrier network — three autonomous systems under performance-based routing, approximately 6,000 portable IPv4 addresses, diverse fiber and transferable transit — under a finished two-story technical headquarters at the edge of Georgia Tech. One title. One closing. Traffic on day one.
2 h flight to 80% of the U.S. population 3 autonomous systems convey ~6,000 IPv4 addresses convey Live production traffic today
The case
In a colo you are a tenant inside someone else’s building, on their power, their network, their security posture, their rules and their next price increase. Every constraint that matters to an AI company — rack density, cooling method, network topology, egress cost, who is allowed through the door — is a conversation with a landlord.
1000 Hemphill inverts that. The building, the land, the electrical service, the fiber, the routing policy and the address are all yours. You set the density. You set the cooling. You set the access list. You announce your own prefixes. Nobody meters your egress, throttles your growth or reprices you at renewal, because there is no renewal.
Reach
Atlanta is not a compromise location. It is the most connected point in the southeastern United States by air and by fiber at once — which is exactly what an inference business needs, because your customers reach you over glass and your enterprise buyers reach you over Delta. Distance is the one input no amount of engineering can undo, and from here it is already short.
80%
of the U.S. population lives within a two-hour flight of Hartsfield–Jackson, the world’s busiest airport, roughly twenty minutes from the building’s front door.
Source: Hartsfield–Jackson Atlanta International Airport fact sheet, reported by CNBC and Airports Council International. Independently verifiable.
Prefill, decode, delivery
A response is three costs stacked end to end: the prefill that produces the first token, the decode loop that produces every token after it, and the network that carries all of them to a human. Rent your infrastructure and you can tune two of the three, then pay a metered bill for the one you cannot touch. Own the building and all three are engineering problems you control.
Time to first token is bound by memory bandwidth and interconnect while the prompt is processed in parallel. On rented instances your batch size, KV-cache policy, tensor split and NIC topology are constrained by whatever instance family had stock that week.
On hardware you own in a building you own, every one of those is a decision your own engineers make on a Tuesday afternoon — not a quota request against a shortage.
Decode is not one long job. It is a long sequence of tiny, jitter-sensitive steps, and the number a user actually perceives is inter-token latency. Scheduler behavior, cache locality and how cleanly tokens leave the box all land directly in that number.
Here, the path from GPU to transit port is a few meters of your own fiber and your own switching, not a virtualized overlay shared with a neighbor you will never meet.
The third cost is the one no GPU benchmark shows you. Your tokens still have to cross a network. Here you own the autonomous systems, the address space and the routing policy — and that policy is chosen by measured latency and loss, not by AS-path length.
Traffic leaves on your own announcements, from the Southeast’s primary interconnection market, with no per-gigabyte meter standing between your model and your user.
A chat turn is one request. An agent turn is a chain: plan, retrieve, rerank, call a tool, call a model, check the result, call again. Each link pays the network cost twice. Twenty milliseconds of avoidable transit, multiplied across a twelve-step loop, is nearly half a second the user feels on every single turn — before your model has done anything wrong.
perceived = (queue + prefill) + n × inter‑token + 2 × steps × RTT
the third term is the one you cannot fix by buying a faster GPU
That is the whole argument for owning the path. Model quality is a race everyone is running. Removing structural latency from your own network is a one-time purchase that compounds on every request you will ever serve.
Standard BGP picks a route by AS-path length and local policy. It has no idea what the latency, loss or jitter of that path actually is, so the route your packets take is the administratively shortest one, not the fastest one. Most networks simply live with the difference.
Two of the three autonomous systems here run a Noction Intelligent Routing Platform deployment. It watches the prefixes that actually carry your traffic, probes them continuously across every available transit path, and injects more-specific routes so outbound traffic leaves on the best-measured provider rather than the default one. It reacts to congestion and brownouts in minutes, not at the next maintenance window.
Outbound is precisely where inference lives. Your tokens are egress. A network whose egress policy is chosen by measurement rather than by AS-path length is a structural advantage that conveys with the building — and it is not something a colocation tenant can buy, because in a cage you take the building’s blended transit and whatever its best-path selection hands you.
Chosen on topology. Performance is never measured.
Chosen on measurement, re-evaluated continuously, enforced by route injection.
Why this facility
The difficult, slow, unglamorous work of a data center is not racking servers. It is power, fiber, address space and an operating record — all four of which exist here, commissioned, and carrying live traffic.
your hardware, your room, your keys
Dedicated utility service with UPS, standby generation and on-site fuel, commissioned and carrying load. Transformer and switchgear lead times are somebody else’s past problem.
Diverse carrier paths along Hemphill with direct transport to Atlanta’s carrier hotels — the densest interconnection point between Ashburn, Miami, Chicago and Dallas.
Three ASNs and a portable IPv4 block convey, two of them already under Noction performance routing. Clean, routed space with established reputation — not a waitlist entry and not a lease from a broker.
Access logs, camera retention, environmental telemetry and change history already accumulating. Every audit you will ever face asks for elapsed time, and this building has it.
The economics
Every month a startup spends on cloud compute buys exactly one month of cloud compute. The cheapest accelerator in the industry is the one you already paid for, running at three in the morning, inside a building nobody can reprice.
Owned capacity has no reservation window, no regional shortage and no instance family being quietly deprecated under your production workload. Your fleet is available because it is yours.
Off-peak hours you already paid for turn into fine-tuning runs, evaluation sweeps and batch inference, instead of a line item with nothing behind it.
You own the ASNs, the address space, the routing policy and the transit relationships. Bandwidth is a contract you negotiate, not a per-gigabyte toll on your own output.
Electricity arrives on your own service at Georgia commercial rates, without a hyperscaler margin stacked on top of every kilowatt-hour your models burn.
Land, building and plant are hard assets with collateral value and a depreciation schedule. Infrastructure stops being pure burn and starts being something a lender recognizes.
Model weights, customer data and training corpora sit on hardware you own, behind a door you control, in a jurisdiction you chose. That sentence closes enterprise deals.
Mix vendors, generations and CPU inference however your kernels want. No instance catalog, no approved-hardware list, no waiting for a provider to stock the part you wrote your code against.
Floor, power, cooling and network are in place today. The distance between signing and serving your first request is how long it takes to rack hardware, not how long it takes to build a facility.
Compute, storage and egress all scale up with success, on prices set by somebody else.
The building does not cost more because you grew. Marginal cost falls as utilization rises.
Certification and sovereignty
Enterprise, healthcare, financial and government buyers do not ask whether your model is good. They ask for a report. The single most expensive line in producing one is proving physical and environmental control over infrastructure you do not own.
In shared space, the physical control families of every framework arrive as inherited controls. You spend the audit reconciling a provider’s report against your own scope, defending subservice-organization carve-outs, mapping a shared-responsibility matrix, and explaining an access list you do not govern. When the landlord changes a procedure, your evidence changes with it, and you find out afterwards.
Here, those controls are first-party. The access list is yours. The cameras and their retention are yours. The visitor log, the environmental monitoring, the suppression test records, the badge revocation process, the media destruction procedure and the change history are all produced by systems you operate, in a building where you are the only tenant. Scope becomes a property line instead of a negotiation, and the hardest evidence in the binder becomes the easiest.
Certification is awarded to an operator’s program, never to a building, and nothing above is offered as a representation that any particular certification is currently held by the property. What an owner-occupied single-tenant facility does is remove the dependency that makes each of these programs slow and expensive: proving control over physical infrastructure that belongs to someone else. Current certifications, audit history and the full controls inventory are disclosed under NDA.
The holding
Engineering, title and network schedules are released to qualified principals under NDA. Figures below are indicative and confirmed in diligence.
Who this is for
Serve tokens from owned accelerators on an owned network, with no egress meter and no instance-family roulette between you and your customers.
Multi-step loops punish network latency more than anything else in the stack. This is the cheapest place to delete it permanently.
Single-tenant envelope, biometric access, first-party evidence. The physical half of every audit stops being the hard half.
An owned, connected, operating U.S. footprint with a real address, a real network and a real engineering office, available now rather than in 2029.
Bandwidth-heavy delivery belongs on your own carrier-dense network. The building has been doing exactly this at scale for years.
Private compute minutes from Atlanta’s carrier hotels, under physical and network control you never have to explain to a landlord.
A training and evaluation cluster at the doorstep of one of the largest engineering schools in the country, with the talent pipeline walking distance away.
Everything a new compute provider needs to bill a customer on day one: floor, power, cooling, network, address space and an operating history.
Build or buy
Recreating this means running every step on the right, in order, with the slowest ones outside your control: a utility interconnection queue, a carrier provisioning cycle, an address-space market and an audit observation window that only passes in real time. Two to three years before a single token ships.
Every one of those steps is already finished here, and the meter on the last one has been running for years. That is what the price reflects — what it replaces, not a multiple of rent.
Continuity and income
Tulix Systems runs live production traffic here today and will execute a long-term leaseback at closing for a secured area of roughly ten cabinets. That does three things at once: it proves the facility performs under real load, it keeps experienced hands in the room through your transition, and it puts income on the asset from day one.
The balance of the building, the offices, the power, the network and both parcels stay open for the buyer. A seller who takes a lease back is a seller telling you what they think of the building.
Guidance
Priced against what it replaces, not against what it rents for.
The figure reflects an operating, carrier-connected, owner-occupied facility with transferable network assets and an address that cannot be assembled on the open market at this location on any timeline a venture-backed company can survive. It is not a multiple of in-place rent and it is not a price per square foot.
Offers are invited. Guidance, the offering memorandum, engineering and title documentation, network and IP schedules, the leaseback structure, floor plans and private site tours are released to qualified principals and their advisors under NDA.
Diligence, answered in advance
Answered here rather than on call three, because none of the answers are uncomfortable.
Tulix Systems is an operating streaming and CDN business, not a property holder. The facility and its network assets are worth considerably more to an operator who will fill them than to a company using a fraction of the floor.
The seller is not exiting the building. Tulix takes a long-term leaseback at closing and remains available for transition services, which is a materially different signal from a seller walking away with the keys.
Yes, and the path is well documented. Autonomous system numbers and IPv4 blocks move through ARIN’s published transfer processes — a merger-and-acquisition transfer where operating assets convey with the numbers, or a specified-recipient transfer where the structure calls for it.
Documentation is prepared during diligence and completion is structured as a closing condition, with officer attestations and supporting records in the form ARIN accepts. The full registry history is in the network schedule released under NDA.
Most carrier master service agreements, transport circuits and cross-connect orders are assignable with the carrier’s consent. Consents are requested early in diligence and tracked as closing conditions alongside title and survey.
Assignment consent is the longest pole in a transaction of this type, which is why the schedule and the consent status are disclosed up front rather than discovered late.
Commissioned electrical capacity, current load, available headroom and the documented upgrade path with the serving utility are all in the engineering package. Those are numbers for diligence, not for a marketing page — but they are numbers we hand over early.
The structural advantage is that density is your decision. There is no landlord, no shared riser, no co-tenant heat load and no building rule between you and rear-door heat exchangers, in-row cooling or a liquid loop. You own the electrical room, the roof and the yard, which is where every one of those upgrades physically lands.
It is operating. Production traffic crosses the building today over live carrier circuits, on conditioned power, behind commissioned security and suppression systems, with staff who work in the offices above it.
You are welcome to stand in the room while it does that. A site visit is arranged for any qualified principal under NDA.
Certification attaches to your program, not to a building — any seller who tells you otherwise is overselling. The honest version is narrower and more valuable: owning a single-tenant facility removes the most expensive dependency in each of those programs.
In shared space the physical and environmental control families arrive as inherited controls, and you spend the engagement reconciling a provider’s report to your scope and defending carve-outs. Here the access list, camera retention, visitor log, environmental telemetry, suppression test records and change history are first-party evidence from systems you run. Scope becomes a property line.
Three things, and it is worth being precise about which is which. First, geography: Atlanta is the southeastern interconnection point and sits on the Florida-to-Northeast fiber corridor, so the physical distance to most U.S. eyeballs is short. Second, path length: transport goes directly to the carrier hotels rather than through an intermediate provider’s aggregation.
Third, and the one most operators skip: routing policy. Two of the three autonomous systems run a Noction Intelligent Routing Platform deployment, which probes the prefixes carrying real traffic across every transit path and injects more-specific routes so outbound traffic uses the best-measured provider instead of the shortest AS-path. Default BGP does not measure performance at all.
What that adds up to is a network tuned on evidence rather than on topology, in a market with short physical paths. We are not going to tell you it is faster than every data center on earth — nobody can prove that, and physics does not care who owns the building. What we will do is give you looking-glass access during diligence so you can measure it yourself against whatever you are running today.
No. Tulix can provide a transition services agreement covering BGP and routing policy, peering relationships, DDoS mitigation, NOC coverage and hands-and-eyes, for as long as you want it — which for most buyers means long enough to hire.
Buyers who would rather run it themselves get documented routing policy, runbooks and a clean handover instead. Both paths are priced in diligence.
Atlanta is the primary interconnection point for the southeastern United States, sitting on the main fiber corridor between Florida and the Northeast. The nearest comparable concentrations of network density are in Ashburn, Miami, Chicago and Dallas — all hundreds of miles away.
Against Northern Virginia specifically you trade a small amount of raw density for the absence of the land, power and interconnection-queue constraints that market is now well known for, at lower cost of power and labor, beside one of the largest engineering schools in the country, two hours’ flight from 80% of the U.S. population. It is also inland: no coastal surge exposure, low seismic risk, no wildfire interface.
Guidance is released to qualified principals and their advisors under NDA, with the offering memorandum, engineering and title documentation, the network and IP schedule and the leaseback structure.
The documentation pack is assembled and ready, so the pace is set by your diligence and by carrier and registry consents rather than by our preparation. Sellers who are ready move quickly; we are ready.
Private diligence
Engineering and title documentation, network and IP schedules, the leaseback structure, floor plans and private site tours are available to qualified principals and their advisors under NDA. Introductions from counsel, brokers and banks are welcome.
All figures, specifications, capacities and structures stated on this page are indicative and subject to buyer diligence, title and engineering verification, carrier and registry assignment consents, and definitive purchase documentation. Nothing here is a representation that any certification is presently held by the property or its operator except where expressly stated. This page is a summary prepared for discussion among qualified principals and does not constitute an offer to sell, a solicitation of an offer to buy, or investment advice.